The rapid expansion of materials databases offers unprecedented opportunities for accelerating materials discovery via machine learning. However, the widespread assumption that larger datasets inherently produce better models does not hold in practice. We propose FUSION (Fusing Uncertainty with Structural Information for Optimal Neural training), an offline dataset pruning strategy that synergistically combines uncertainty quantification with crystallographic structure analysis via geometric fingerprinting, framing dataset pruning as a discrete optimization problem. Through evaluation across 3 benchmark datasets, FUSION consistently outperforms baselines, including random pruning, uncertainty sampling, weighting factor pruning, diversity sampling, and active learning. It demonstrates robust transferability across 11 diverse architectures, outperforming random pruning by 1.91–13.65% across different datasets, with an average improvement of 6.36%. Moreover, our analysis suggests that different models exhibit varying robustness characteristics when faced with pruned training data, highlighting the importance of model selection tailored to dataset composition. We identify optimal pruning points where removing just 0–8% of training data improves model performance, yielding gains up to 12.67% in specific model–dataset combinations. These results establish a new paradigm for materials informatics that prioritizes data quality over quantity, offering a pathway toward more efficient and sustainable machine learning workflows in computational materials science.

With the rapid proliferation of large language models (LLMs), model pools have become increasingly heterogeneous in both capability and efficiency.Larger LLMs can improve quality but incur higher latency and cost, while smaller LLMs are the opposite, making perquery model selection crucial in practice.This has spawned LLM routers that dispatch each query to an appropriate model.Existing routers lack fine-grained resource awareness across deployment settings, which degrades efficiency metrics in real-world serving.To this end, we propose FLARE, a length-centric, resourceaware multi-LLM routing framework that uses length-based models to estimate per-query latency and cost.FLARE formulates routing as a discrete multi-objective optimization problem to achieve an efficient trade-off.Experiments show that FLARE reduces latency and cost by up to 68% and 75% while achieving sufficient accuracy, and can be easily applied to new datasets and LLMs.

Modern large language model (LLM) serving systems confront inefficient GPU utilization due to the fundamental mismatch between compute-intensive prefill phase and memory-bound decode phase. While current practices attempt to address this by organizing these phases into hybrid batches, such solutions create an inefficient tradeoff that sacrifices either throughput or latency, leaving substantial GPU resources underutilized. For this, we identify two key root causes: 1) the prefill phase suffers from suboptimal compute utilization due to wave quantization and attention bottlenecks, and 2) hybrid batching disproportionately prioritizes latency over throughput, wasting both compute resources and memory bandwidth. To mitigate the issues, we present Bullet, a novel spatial-temporal orchestration system that eliminates these inefficiencies through fine-grained phase coordination. Bullet enables concurrent execution of prefill and decode requests, while dynamically provisioning GPU resources based on real-time performance modeling. By integrating SLO-aware scheduling and adaptive resource allocation, Bullet maximizes GPU utilization without compromising latency targets. Experimental evaluations on real-world workloads demonstrate that Bullet delivers 1.26× average throughput gains (up to 1.55×) over state-of-the-arts, while consistently meeting latency constraints.

Particle-in-Cell (PIC) simulations devote most cycles to particle-grid interactions, and their fine-grained atomic updates become a severe bottleneck on traditional many-core CPUs. The evolution of CPU architectures, particularly the integration of specialized Matrix Processing Units (MPUs) designed for efficient matrix outer-product operations, presents a paradigm shift and an opportunity to alleviate these bottlenecks. Capitalizing on this architectural advancement, this work focuses on adapting the critical current deposition step in PIC simulations to this new matrix-centric computational model.

Spatial accelerators are promising to satiate the growing demands for performance and energy efficiency in deep neural networks (DNNs). Due to the speed gap between onchip compute cores and off-chip memory bandwidth, common DNNs suffer from poor operational intensity and are increasingly memory-bound. While operator fusion has shown potential in alleviating this bottleneck, existing approaches suffer from two key limitations. They rely on predefined fusion templates before tensor mapping and impose tile constraints during mapping. As a result, they overlook the potential of fusing more operators and lead to sub-optimal performance. In this paper, we propose a segment fusion dataflow optimization framework called SFD. Central to this framework is the dataflow abstraction that enables template-free operator fusion after mapping and supports tile constraint relaxation through tile scheduling. Based on this abstraction, we first introduce a memory-centric mapper, which defines a design space and incorporates an algorithm to facilitate design space exploration (DSE). Then we propose an analytical network segmenter, which leverages mapping results to analyze tensor lifetimes and on-chip memory usage, fusing operators into variable-length segments. Finally, we introduce a dependency-aware tile scheduler, which develops a priority queue for each segment to ensure correct execution order. Extensive experiments with different DNNs demonstrate SFD achieves$1.4 \times$to$2.2 \times$speedup for spatial accelerators over state-of-the-art fusion frameworks.

Particle-in-Cell (PIC) simulations are fundamental to plasma physics but often suffer from limited scalability due to particle–grid interaction bottlenecks and particle redistribution costs. Specifically, the particle–grid interaction computations have not taken full advantage of the emerging Matrix Processing Units (MPUs), the particle motion introduces irregular memory accesses, and the bulk-synchronous redistribution further destroys long-term data locality thereby limiting parallel efficiency. To address these inefficiencies, we present POLAR-PIC, a co-designed framework for large-scale PIC simulations that (i) reformulates Field Interpolation into an MPU-friendly outer-product form, (ii) maintains a physically ordered particle layout to preserve memory contiguity, and (iii) overlaps particle communication with Deposition to hide redistribution overhead. The evaluation on the pilot system of an Exascale supercomputer demonstrates that POLAR-PIC accelerates the entire particle-processing phase by up to 10.9 × in uniform plasma and 4.4 × in real-world laser-ion acceleration scenarios compared to the native WarpX reference pipeline on LX2. Ablation studies reveal that the speedups achieved by Interpolation and Deposition are 8.0 × and 13.2 × , respectively, and the asynchronous communication design sustains a \(99.1\%\) overlap ratio. In cross-platform comparisons, POLAR-PIC achieves \(13.2\%\) of theoretical peak efficiency on the CPU-based LS system, while WarpX reaches \(9.6\%\) on NVIDIA A800 GPUs. Notably, the scalability evaluation demonstrates that POLAR-PIC maintains \(67.5\%\) weak scaling efficiency on over 2 million cores under high-migration dynamic workloads, highlighting the importance of holistic co-design for future matrix-centric HPC systems.

Large-scale simulation of Spiking Neural Networks (SNNs) on supercomputers is pivotal for unraveling the mechanisms of brain function and advancing brain-inspired intelligence. However, efficiently mapping billions of neurons onto distributed nodes presents a significant challenge due to the heterogeneity of neuronal activities and complex, irregular network connectivity. While static partitioning strategies perform well in stable states, they often fail under metastable neurodynamics where theoretical models cannot accurately predict neuronal firing rates, leading to severe load imbalance. To address this, we propose TADS (Trend-Aware Dynamic load balancing with Delay-Sharded graph infrastructure), a framework tailored for large-scale SNN simulations. First, we identify the specific failure modes of static partitioning under metastable dynamics, establishing the necessity for runtime intervention. Second, we introduce a trend-aware dynamic load balancing strategy. By analyzing the temporal evolution of loads, this approach effectively distinguishes persistent imbalance from transient fluctuations, thereby avoiding unnecessary migrations caused by momentary jitter. Third, we design a delay-sharded graph infrastructure that leverages synaptic delays to parallelize graph modifications, significantly reducing the overhead associated with dynamic load balancing. Experimental results on the Tianhe-Xingyi supercomputer demonstrate that TADS effectively handles metastable scenarios, achieving up to 2.13 × speedup over state-of-the-art static partitioning methods, while sustaining 1.58 × performance improvement at the largest evaluated scale of 192 nodes.

Kolmogorov–Arnold Networks (KANs) promise higher expressive capability and stronger interpretability than Multilayer Perceptron, particularly in the domain of AI for Science. However, practical adoption has been hindered by low GPU utilization of existing parallel implementations. To address this challenge, we present a GPU-accelerated operator library, named PolyKAN, which is the first general open-source implementation of KAN and its variants. PolyKAN fuses the forward and backward passes of polynomial KAN layers into a concise set of optimized CUDA kernels. Four orthogonal techniques underpin the design: (i) lookup-table with linear interpolation that replaces runtime expensive math-library functions; (ii) 2D tiling to expose thread-level parallelism with preserving memory locality; (iii) a two-stage reduction scheme converting scattered atomic updates into a single controllable merge step; and (iv) coefficient-layout reordering yielding unit-stride reads under the tiled schedule. PolyKAN can deliver 1.2–10 × faster inference and 1.4–12 × faster training than a Triton + cuBLAS baseline, with identical accuracy on speech, audio-enhancement, and tabular-regression workloads on both highend GPU and consumer-grade GPU.

Sparse Matrix–Matrix Multiplication (SpMM) is a core kernel in scientific computing, data analytics, and artificial intelligence, supporting applications such as linear solvers and Graph Neural Networks (GNNs). The Scalable Matrix Extension (SME) in Armv9 introduces dedicated matrix acceleration for ARM CPUs, but exploiting its full potential for SpMM requires architecture-aware optimizations to address irregular sparsity and hardware constraints.

Generalized Category Discovery (GCD) is a challenging task that aims to identify both seen and novel categories in unlabeled data. We argue that a clear margin between seen and novel class representations is essential for accurate recognition. However, existing methods often ignore this margin, mapping representations to prototypes without enforcing separation between seen and novel classes. This leads to a bias where seen samples are misclassified as novel. To address this issue, we propose DebiasGCD, a debiasing framework that enhances prototype separation through margin-aware learning. Unlike prior work that relies on static prototype learning and overlooks fine-grained representations, our method introduces Dynamic Prototype Debiasing (DPD) and Spatial-Aware Representation Distillation (SARD) to mitigate this bias. First, DPD dynamically enforces inter-prototype margins, improving class-specific feature learning and prototype discrimination. Meanwhile, SARD promotes local representation of spatial learning, supporting DPD to capture subtle details that further refine class-specific features. By synergizing these components, DebiasGCD significantly improves prototype discriminability, generating more reliable predictions for seen classes. Extensive experiments demonstrate that our approach effectively mitigates pseudo-labeling bias across datasets, especially on fine-grained ones, achieving +8.3% and +9.6% improvements on the ‘All’ classes in CUB and Stanford Cars, respectively.

The rapid growth of model parameters presents a significant challenge when deploying large generative models on GPU. Existing LLM runtime memory management solutions tend to maximize batch size to saturate GPU device utilization. Nevertheless, this practice leads to situations where the KV Cache of certain sequences cannot be accommodated on GPUs with limited memory capacity during the model inference, requiring temporary eviction from GPU memory (referred to as KV Cache spillover). However, without careful consideration of the LLM inference's runtime pattern, current LLM inference memory management solutions face issues like one-size-fits-all spillover handling approach for different platforms, under-utilization of GPU in prefill stage, and suboptimal sequence selection due to direct employment of swap or recomputation. In this paper, we introduce FuseSpill, a holistic KV Cache management solution designed to boost LLM inference on memory-constrained GPU by efficiently handling KV Cache spillover. Specifically, FuseSpill consists of a spillover cost model that analyzes the system cost of spillover handling techniques quantitatively, a KV cache swap orchestrator to further refine the basic swap technique to sophisticated disaggregate KV Cache across heterogeneous devices for decoding iterations, a multi-executor scheduler to effectively coordinate task executors across devices, and a response length predictor to exploit the length-aware sequence selection strategy when KV Cache spillover occurs. The experimental results demonstrate that our implementation outperforms existing solutions, delivering a 20% to 40% increase in throughput while simultaneously reducing the inference latency of the spillover sequences.

LLM deployment on end-user devices has attracted significant interest from tech giants and research institutions due to privacy benefits and the elimination of network roundtrips. Reducing latency is crucial for improving user experience. Enduser devices, such as desktop and mobile processors, often integrate CPU and GPU on a single die, making tensor parallelism a promising approach to distribute workloads and reduce inference latency. However, the predefined tensor parallelism in traditional practices cannot guarantee optimization due to heterogeneity and resource contention in the CPU-GPU integrated end-user devices. In this paper, we propose Doppeladler, a practical framework designed to facilitate parallel inference of LLM on end-user devices. Doppeladler adaptively optimizes tensor parallelism based on real-time conditions and device status, enhancing resource utilization by refining workload partitioning and scheduling with a heuristic-based approach. Additionally, it lowers the costs of CPU-GPU hybrid execution by managing device resource, trigger workload rebalance during runtime to mitigate the penalty of contention, and minimizing communication overhead through zero-copy techniques and alternate access pattern. Experiments demonstrate that Doppeladler outperforms existing methods by 1.2 to 2.5 times.

GPUs have been heavily utilized in diverse applications, and numerous approaches, including kernel fusion, have been proposed to boost GPU efficiency through concurrent kernel execution. However, these approaches generally overlook the opportunities to mitigate warp stalls and improve instruction level parallelism (ILP) in inter-kernel resource sharing. To address this issue, we introduce GOPTX, a novel design for kernel fusion that improves ILP through deliberate weaving instructions at the PTX level. GOPTX establishes a merged control flow graph (CFG) from original kernels, enabling to interleaving of instructions that were sequentially executed by default and minimizing pipeline stalls on data hazards. We further propose a latency-aware instruction weaving algorithm for more efficient instruction scheduling and an adaptive code slicing method to enlarge the scheduling space. Experimental evaluation demonstrates that GOPTX achieves an average speedup of $\mathbf{1 1. 2 \%}$ over the baseline concurrent execution, with a maximum improvement of 23%. The hardware resource utilization statistics show significant enhancements in eligible warps per cycle and resource use.

LoRA is a widely used Parameter-Efficient FineTuning (PEFT) technique for customizing pre-trained backbone models to specific tasks. Serving a backbone model with numerous LoRA adapters, known as multi-tenant LoRA serving, is a common scenario where different users utilize distinct LoRA adapters while sharing the same backbone model. To support more LoRA adapters simultaneously and improve efficiency, existing solutions dynamically load adapters from host memory and separate workloads into batched backbone model computation and adapter computation. However, they introduce complex data dependencies and necessitate careful coordination of loading and computation to enhance efficiency. We introduce AuLoRA, a multi-tenant LoRA serving system that achieves fine-grained orchestration of adapter loading and computation alongside backbone model execution. It optimizes both Time-to-First-Token (TTFT) and throughput by: 1) layer-wise-priority LoRA adapter loading, which reorganizes adapter loading by layer, to perform inference before adapters are fully loaded and overlaps adapter loading with backbone computation. 2) Intra-layer pipelined LoRA adapter execution, which loads LoRA adapters and performs computation in a pipelined manner, further hiding the adapter loading overhead. 3) Dynamic LoRA adapter batching, which explores the optimal LoRA adapter batching plan by comprehensively considering kernel launch overhead and modern hardware parallelism, improving computational efficiency. We compare AuLoRA with S-LoRA, a state-of-the-art multi-tenant LoRA serving system, and the results show that AuLoRA can achieve up to$3.03 \times$TTFT reduction and$1.27 \times$throughput improvement.

Deep Neural Network (DNN) inference has become a cornerstone of latency-sensitive applications such as autonomous driving and augmented reality. While GPUs offer high throughput for DNN inference, they often suffer from underutilization due to coarse-grained scheduling and limited concurrency. Existing GPU-sharing methods either lack awareness of fine-grained kernel interference or fail to meet service-level objectives (SLOs) under multi-tenant, multi-priority workloads. In this paper, we propose IasRT, a runtime framework that enables interferenceaware and SLO-driven GPU sharing for real-time DNN inference. IasRT profiles kernel-level resource usage and interference sensitivity, and dynamically partitions GPU streaming multiprocessors (SMs) to collocate jobs with minimal performance degradation. Furthermore, it introduces a dynamic SLO controller to maintain latency targets for multiple latency-sensitive (LS) jobs simultaneously. Evaluations on real-world DNN workloads show that IasRT reduces the 99th percentile latency of LS jobs by up to 38% compared to the state-of-the-art GPU sharing methods, while maintaining similar overall throughput from multiple collocated workloads, demonstrating its effectiveness in high-concurrency environments.

Supervised machine learning techniques are increasingly being adopted to speed up electronic structure predictions, serving as alternatives to first-principles methods like Density Functional Theory (DFT). Although current DFT datasets mainly emphasize chemical properties and atomic forces, the precise prediction of electronic charge density is essential for accurately determining a system's total energy and ground state properties. In this study, we introduce a novel electronic charge density dataset named ECD, which encompasses 140,646 stable crystal geometries with medium-precision Perdew–Burke–Ernzerhof (PBE) functional data. Within this dataset, a subset of 7,147 geometries includes high-precision electronic charge density data calculated using the Heyd–Scuseria–Ernzerhof (HSE) functional in DFT. By designing various benchmark tasks for crystalline materials and emphasizing training with large-scale PBE data while fine-tuning with a smaller subset of high-precision HSE data, we demonstrate the efficacy of current machine learning models in predicting electronic charge densities. The ECD dataset and baseline models are open-sourced to support community efforts in developing new methodologies and accelerating materials design and applications.

Due to the rise in antimicrobial resistance, identifying novel compounds with antibiotic potential is crucial for combatting this global health issue. However, traditional drug development methods are costly and inefficient. Recognizing the pressing need for more effective solutions, researchers have turned to machine learning techniques to streamline the prediction and development of novel antibiotic compounds. While foundation models have shown promise in antibiotic discovery, current mainstream efforts still fall short of fully leveraging the potential of multimodal molecular data. Recent studies suggest that contrastive learning frameworks utilizing multimodal data exhibit excellent performance in representation learning across various domains. Building upon this, we introduce CL-MFAP, an unsupervised contrastive learning (CL)-based multimodal foundation (MF) model specifically tailored for discovering small molecules with potential antibiotic properties (AP) using three types of molecular data. This model employs 1.6 million bioactive molecules with drug-like properties from the ChEMBL dataset to jointly pretrain three encoders: (1) a transformer-based encoder with rotary position embedding for processing SMILES strings; (2) another transformer-based encoder, incorporating a novel bi-level routing attention mechanism to handle molecular graph representations; and (3) a Morgan fingerprint encoder using a multilayer perceptron, to achieve the contrastive learning purpose. The CL-MFAP outperforms baseline models in antibiotic property prediction by effectively utilizing different molecular modalities and demonstrates superior domain-specific performance when fine-tuned for antibiotic-related property prediction tasks.

As the model size continuously increases, pipeline parallelism shows great promise in throughput-oriented LLM inference due to its low demand on communications. However, imbalanced pipeline workloads and complex data dependencies in the prefill and decode phases result in massive pipeline bubbles and further severe performance reduction.

In this paper, we introduce CFDONEval, a comprehensive evaluation of 12 operator-learning-based neural network (ON) models to simulate 7 benchmark fluid dynamics problems. These problems cover a range of 2D scenarios, including Darcy flow, two-phase flow, Taylor-Green vortex, lid-driven cavity flow, tube flow, circular cylinder flow, and 3D periodic hill flow. For a rigorous evaluation, we establish 22 fluid dynamics datasets for these benchmark problems, 18 of which are newly generated using traditional numerical methods, such as the finite element method. Our evaluation tackles 5 key challenges: multiscale phenomena, convection dominance, long-term predictions, multiphase flows, and unstructured meshes over complex geometries. We assess computational accuracy, efficiency, and flow field visualization, offering valuable insights into the application of ON models in fluid dynamics research. Our findings show that attention-based models perform well in handling almost all challenges; models with a U-shaped structure excel in handling multiscale problems; and the NU-FNO model demonstrates the smallest relative error in L2 norm when processing nonuniform grid data. The related code, dataset, and appendix are publicly available at: https://github.com/Sysuzqs/CFDNNEval.

Pipeline parallelism has emerged as a predominant approach for deploying large language models (LLMs) across distributed nodes, owing to its lower communication overhead compared to tensor parallelism. While demonstrating high throughput in request serving, pipeline parallelism often faces performance limitations caused by pipeline bubbles, which are primarily resulted from imbalanced computation delays across batches. Existing methods like Sarathi-Serve attempt to address this through hybrid scheduling of chunked prefill and decode tokens with a fixed token budget. However, such methods may still experience significant fluctuations, arising either from insufficient prefill tokens or uneven distribution of decode tokens, ultimately leading to computational imbalance.

The increasing interconnectivity of HPC systems has highlighted the need for efficient application migration across different environments. Containers, widely adopted for this purpose, simplify deployment but often fail to deliver optimal performance due to the separated build and execution container workflow. This leads to generic container images that miss out on system-specific software stack advantages, a challenge we define as the adaptability issue.

Stencil computations are fundamental to various HPC and intelligent computing applications, often consuming significant execution time. The emergence of specialized matrix units presents new opportunities to accelerate stencil computations. While scalable matrix compute units provide substantial computing horsepower, prior efforts fail to fully utilize the computing capabilities for stencils due to suboptimal matrix-unit utilization, limited instruction-level parallelism, and low cache hit rates. This paper introduces HStencil, a novel stencil computing framework utilizing matrix and vector units. HStencil addresses these challenges through three contributions: 1) microkernels that jointly leverage matrix and vector units to enhance hardware utilization; 2) fine-grained instruction scheduling with interleaved execution to enhance instruction-level parallelism; and 3) spatial prefetch to sustain high performance when working sets exceed cache capacity. Evaluations on representative benchmarks demonstrate that HStencil achieves maximum speedups of 1.81x – 5.76x over auto-vectorization across different CPU platforms, delivers 31% - 91% higher performance versus state-of-the-art methods.

This paper proposes a unified optimization framework for a power splitting (PS)-based multicarrier multiple-input and multiple-output (MIMO) simultaneous wireless information and power transfer (SWIPT) system with Tomlinson-Harashima pre coding (THP)-based non-linear transceivers or linear transceivers. Our aim is to minimize the transmit power under the sum mean-square-error (MSE) and energy harvesting (EH) constraints. To solve this formulated non-convex problem, we propose a structural solution (SS) which applies the closed-form expressions of equalization matrices, feedback matrices and precoding matrices to establish an equivalent optimization problem in terms of the power allocation and PS ratio. Then the equivalent problem is solved by a two-layer optimization scheme. Simulation results show that the THP-based non-linear transceivers need less transmit power than linear transceivers to achieve the same performance of EH and sum MSE.

Spatial accelerators are domain-specific architectures to elevate performance and energy efficiency for deep neural networks (DNNs). They also bring a large number of schedule parameters to determine computation and data movement patterns of DNNs. Previous works formulate the schedule problem as design space exploration or integer linear programming. However, these advanced techniques face the challenge of efficiency or scalability. In this article, we propose Ceiba, which is a deep reinforcement learning-based DNN scheduler for spatial accelerators. Ceiba observes the running DNN computation as well as the spatial architecture to make schedule decisions. Then, Ceiba receives a reward to learn and produce the best-fit policy. To provide efficient and scalable scheduling, Ceiba constructs a DNN-architecture-specific action space. It is defined by upper and lower bounds to exclude invalid and sub-optimal schedule candidates. Extensive experiments demonstrate that Ceiba generally provides better performance for spatial accelerators under a fixed number of searching steps or a fixed amount of time. Specifically, Ceiba achieves an average 2.2× speedup for the Simba accelerator, compared with the state-of-the-art scheduler. When scaling the batch size and the hardware architecture up by 64×, the performance gains of Ceiba are 1.8× and 1.2× on average, respectively. Moreover, Ceiba exhibits better scalability for the Eyeriss accelerator.

Transformer-based models have unlocked a plethora of powerful intelligent applications at the edge, such as voice assistant in smart home. Traditional deployment approaches offload the inference workloads to the remote cloud server, which would induce substantial pressure on the backbone network as well as raise users' privacy concerns. To address that, in-situ inference has been recently recognized for edge intelligence, but it still confronts significant challenges stemming from the conflict between intensive workloads and limited on-device computing resources. In this paper, we leverage our observation that many edge environments usually comprise a rich set of accompanying trusted edge devices with idle resources and proposeGalaxy+, a collaborative edge AI system that breaks the resource walls across heterogeneous edge devices for efficient Transformer inference acceleration.Galaxy+introduces a novel hybrid model parallelism to orchestrate collaborative inference, along with a heterogeneity and memory-aware parallelism planning for fully exploiting the resource potential. To mitigate the impact of tensor synchronizations on inference latency under bandwidth-constrained edge environments,Galaxy+devises a tile-based fine-grained overlapping of communication and computation. Furthermore, a fault-tolerant re-scheduling mechanism is developed to address device-level resource dynamics, ensuring stable and low-latency inference. Extensive evaluation based on prototype implementation demonstrates thatGalaxy+remarkably outperforms state-of-the-art approaches under various edge environment setups, achieving a$1.2\times$to$4.24\times$end-to-end latency reduction. Besides,Galaxy+can adapt to device-level resource dynamics, swiftly rescheduling and restoring inference in the presence of unexpected straggler devices.

Persistent memory (PM) brings important opportunities for improving data storage including the widely used hash tables. However, PM is not friendly to small writes, which causes existing PM hashes to suffer from high hardware write amplification. Hybrid memory offers the performance and concurrency of DRAM and the durability and capacity of PM, but existing hybrid memory hashes cannot deliver high performance, low DRAM footprint, and fast recovery at the same time. This paper proposes WALSH, a flat hash with novel log-structured separate chaining designs to optimize the performance while ensuring low DRAM footprint and fast recovery. To address the overhead of hash resizing and garbage collection (GC), WALSH further proposes partial resizing/GC mechanisms and a 4-phase protocol for concurrent hash operations. As a result, WALSH is the first flat index for hybrid memory with embedded write aggregation ability. A comprehensive evaluation shows that WALSH substantially outperforms state-of-the-art hybrid memory hashes; e.g., its insert throughput is up to 2.4X that of related works while saving more than 87% of DRAM. WALSH also provides efficient recovery; e.g., it can recover a dataset with 1 billion objects in just a few seconds.

Transformer models have shown significant success in a wide range of tasks. However, the massive resources required for its inference prevent deployment on a single device with relatively constrainted resources, thus leaving a high threshold of integrating their advancements. Observing scenarios such as smart home applications on edge devices and cloud deployment on commodity hardware, it is promising to distribute Transformer inference across multiple devices. Unfortunately, due to the tightly-coupled feature of Transformer model, existing model parallelism approaches necessitate frequent communication to resolve data dependencies, making them unacceptable for distributed inference, especially under relatively weak interconnection. In this paper, we propose DeTransformer, a communication-efficient distributed Transformer inference system. The key idea of DeTransformer involves the co-design of Transformer architecture to reduce the communication during distributed inference. In detail, DeTransformer is based on a novel block parallelism approach, which restructures the original Transformer layer with a single block to the decoupled layer with multiple sub-blocks. Thus, it can exploit model parallelism between sub-blocks. Next, DeTransformer contains an adaptive execution approach that strikes a trade-off among communication capability, computing power and memory budget over multiple devices. It incorporates a two-phase planning for execution, namely static planning and runtime planning. The static planning runs offline, containing a profiling procedure and a weight placement strategy before execution. The runtime planning dynamically determines the optimal parallel computing strategy from an expertly crafted search space based on real-time requests. Notably, this execution approach can adapt to heterogeneous devices by distributing workload based on devices’ computing capabilities. We conduct experiments for both auto-regressive and auto-encoder tasks of Transformer models. Experimental results show that DeTransformer can reduce distributed inference latency by up to 2.81× compared to the SOTA approach on 4 devices, while effectively maintaining task accuracy and a consistent model size.

We are thrilled to share the program for this year's IEEE Cluster conference, showcasing a diverse range of topics in cluster computing and emphasizing the field's ongoing significance. The program strikes a balance between established subjects like architectures, software environments, and scientific applications, and emerging areas such as data analytics and deep learning.

Generalized Category Discovery (GCD) aims to identify a mix of known and novel categories within unlabeled data sets, providing a more realistic setting for image recognition. Essentially, GCD needs to remember existing patterns thoroughly to recognize novel categories. Recent state-of-the-art method SimGCD transfers the knowledge from known-class data to the learning of novel classes through debiased learning. However, some patterns are catastrophically forgot during adaptation and thus lead to poor performance in novel categories classification. To address this issue, we propose a novel learning approach, LegoGCD, which is seamlessly integrated into previous methods to enhance the discrimination of novel classes while maintaining performance on previously encountered known classes. Specifically, we design two types of techniques termed as Local Entropy Regularization (LER) and Dual-views Kullback-Leibler divergence constraint (DKL). The LER optimizes the distribution of potential known class samples in unlabeled data, thus ensuring the preservation of knowledge related to known categories while learning novel classes. Meanwhile, DKL introduces Kullback-Leibler divergence to encourage the model to produce a similar prediction distribution of two view samples from the same image. In this way, it successfully avoids mismatched prediction and generates more reliable potential known class samples simultaneously. Extensive experiments validate that the proposed LegoGCD effectively addresses the known category forgetting issue across all datasets, e.g., delivering a 7.74% and 2.51% accuracy boost on known and novel classes in CUB, respectively. Our code is available at: https://github.com/Cliffia123/LegoGCD.

Transformer models have shown significant success in a wide range of tasks. Meanwhile, massive resources required by its inference prevent scenarios with resource-constrained devices from in-situ deployment, leaving a high threshold of integrating its advances. Observing that these scenarios, e.g. smart home of edge computing, are usually comprise a rich set of trusted devices with untapped resources, it is promising to distribute Transformer inference onto multiple devices. However, due to the tightly-coupled feature of Transformer model, existing model parallelism approaches necessitate frequent communication to resolve data dependencies, making them unacceptable for distributed inference, especially under weak interconnect of edge scenarios. In this paper, we propose DeTransformer, a communication-efficient distributed in-situ Transformer inference system for edge scenarios. DeTransformer is based on a novel block parallelism approach, with the key idea of restructuring the original Trans-former layer with a single block to the decoupled layer with multi-ple sub-blocks and exploit model parallelism between sub-blocks. Next, DeTransformer contains an adaptive placement approach to automatically select the optimal placement strategy by striking a trade-off among communication capability, computing power and memory budget. Experimental results show that DeTransformer can reduce distributed inference latency by up to 2.81 x compared to the SOTA approach on 4 devices, while effectively maintaining task accuracy and a consistent model size.

In addressing the challenge of Crystal Structure Prediction (CSP), symmetry-aware deep learning models, particularly diffusion models, have been extensively studied, which treat CSP as a conditional generation task. However, ensuring permutation, rotation, and periodic translation equivariance during diffusion process remains incompletely addressed. In this work, we propose EquiCSP, a novel equivariant diffusion-based generative model. We not only address the overlooked issue of lattice permutation equivariance in existing models, but also develop a unique noising algorithm that rigorously maintains periodic translation equivariance throughout both training and inference processes. Our experiments indicate that EquiCSP significantly surpasses existing models in terms of generating accurate structures and demonstrates faster convergence during the training process.

Transformer-based models have unlocked a plethora of powerful intelligent applications at the edge, such as voice assistant in smart home. Traditional deployment approaches offload the inference workloads to the remote cloud server, which would induce substantial pressure on the backbone network as well as raise users’ privacy concerns. To address that, in-situ inference has been recently recognized for edge intelligence, but it still confronts significant challenges stemming from the conflict between intensive workloads and limited on-device computing resources. In this paper, we leverage our observation that many edge environments usually comprise a rich set of accompanying trusted edge devices with idle resources and propose Galaxy, a collaborative edge AI system that breaks the resource walls across heterogeneous edge devices for efficient Transformer inference acceleration. Galaxy introduces a novel hybrid model parallelism to orchestrate collaborative inference, along with a heterogeneity-aware parallelism planning for fully exploiting the resource potential. Furthermore, Galaxy devises a tile-based fine-grained overlapping of communication and computation to mitigate the impact of tensor synchronizations on inference latency under bandwidth-constrained edge environments. Extensive evaluation based on prototype implementation demonstrates that Galaxy remarkably outperforms state-of-the-art approaches under various edge environment setups, achieving up to 2.5× end-to-end latency reduction.

ARM-based many-core processors in supercomputers enable exascale-level data analysis by leveraging the massive number of cores within a single chip. However, most existing big data frameworks, designed for distributed environments, fail to effectively harness the full potential of advanced supercomputing systems. Two significant issues impeding data processing efficiency are the remote memory access between NUMA nodes and the disregard for architectures within compute nodes. Furthermore, the explosive growth of intermediate data results in a severe mismatch between I/O capabilities and computation performance due to frequent read and write operations. In this paper, we present SparkTH, a NUMA-aware and I/O-efficient framework for big data processing on many-core supercomputing systems. SparkTH incorporates a NUMA resource management layer to fully utilize many-core resources and cuts the number of I/O operations in half in parallel file systems. We evaluated SparkTH on two ARM-based many-core systems, and it outperformed the original Spark by up to 2.1× on typical big data benchmarks and 8.7× on scientific computing applications.

Spiking Neural Network (SNN) is an effective tool for the simulation of neuronal dynamics as well as the understanding of brain structure and functions. However, scaling up SNN for large-scale simulations poses significant computational demands that necessitate the supercomputers. The advent of distributed simulation introduces the requirement of SNN partition and process mapping, which becomes a critical challenge in the context of large-scale distributed SNN simulations. In this paper, we propose an Area-based partition and Topology-aware process Mapping (ATM) strategy to balance the computation workload while coping with the heterogeneity of communication interconnect. We first model the computation workload and communication volume of the SNN simulation according to its biological features. Based on this model, we design an area-based SNN partition strategy to balance the computation workload. Subsequently, we introduce a topology-aware strategy for process mapping, Bottleneck Fulfilling (BF), tailored specifically for collective communication paradigms. Experiments are conducted on an HPC cluster with a multi-area model of the marmoset brain. The results demonstrate that the proposed approach achieves up to 2.2x speedup compared with the baseline on 290 compute nodes.

Featuring mixed-precision tensor operations, accelerators significantly enhance performance for many error-tolerant computing tasks, but their applicability is limited in scenarios demanding high precision. While emulating higher-precision data types from lower-precision ones can bridge this gap, existing techniques either struggle to achieve sufficient accuracy or incur excessive overhead, inevitably negating performance gains. To mitigate the issue, we propose MixPert, a novel system that balances performance and accuracy via optimizing single-precision emulation on GPU Integer Tensor Cores. MixPert devises an efficient data layout and augments the computation pipeline on Tensor Cores. By deeply analyzing performance-precision trade-offs, MixPert provides users with multiple configurations based on accuracy requirements. Furthermore, MixPert can seamlessly integrate with compilers, facilitating automatic adaptation and tuning of mixed-precision parameters. Evaluations on real-world scientific computing and deep learning applications demonstrate that MixPert achieves an average speedup of 1.72× compared to cuBLAS on general-purpose cores. Beyond maintaining improved accuracy, MixPert outperforms state-of-the-art approaches APE and CUTLASS by 1.22× and 1.21×, respectively.

Superconductivity is a fascinating phenomenon observed in certain materials under certain conditions. However, some critical aspects of it, such as the relationship between superconductivity and materials' chemical/structural features, still need to be understood. Recent successes of data-driven approaches in material science strongly inspire researchers to study this relationship with them, but a corresponding dataset is still lacking. Hence, we present a new dataset for data-driven approaches, namely SuperCon3D, containing both 3D crystal structures and experimental superconducting transition temperature (Tc) for the first time. Based on SuperCon3D, we propose two deep learning methods for designing high Tc superconductors. The first is SODNet, a novel equivariant graph attention model for screening known structures, which differs from existing models in incorporating both ordered and disordered geometric content. The second is a diffusion generative model DiffCSP-SC for creating new structures, which enables high Tc-targeted generation. Extensive experiments demonstrate that both our proposed dataset and models are advantageous for designing new high Tc superconducting candidates.

Distributed large model inference is still in a dilemma where balancing cost and effect. The online scenarios demand intraoperator parallelism to achieve low latency and intensive communications makes it costly. Conversely, the inter-operator parallelism can achieve high throughput with much fewer communications, but it fails to enhance the effectiveness.

Direct numerical simulation (DNS) is a technique that directly solves the fluid Navier-Stokes equations with high spatial and temporal resolutions, which has driven much research regarding the nature of turbulence. For high-Reynolds number (Re) incompressible turbulence of particular interest, where the nondimensional Re characterizes the flow regime, the application of DNS is hindered by the fact that the numerical grid size (i.e., the memory requirement) scales with Re3, while the overall computational cost scales with Re4. Recent studies have shown that developing efficient parallel methods for heterogeneous many-core systems is promising to solve this computational challenge.

Remote Memory Access (RMA) enables direct access to remote memory to achieve high performance for HPC applications. However, most modern parallel programming models lack schemes for the remote process to detect the completion of RMA operations. Many previous works have proposed programming models and extensions to notify the communication peer, but they did not solve the multi-NIC aggregation, portability, hardware-software co-design, and usability problems. In this work, we proposed a Unified Notifiable RMA (UNR) library for HPC to address these challenges. In addition, we demonstrate the best practice of utilizing UNR within a real-world scientific application, PowerLLEL. We deployed UNR across four HPC systems, each with a different interconnect. The results show that PowerLLEL powered by UNR achieves up to a 36% acceleration on 1728 nodes of the Tianhe-Xingyi supercomputing system.

Recently, the sparsely-gated Mixture-Of-Experts (MoE) architecture has garnered significant attention. To benefit a wider audience, fine-tuning MoE models on more affordable clusters, which are typically a limited number of bandwidthconstrained GPU nodes, holds promise. However, it is non-trivial to apply existing cost-effective fine-tuning approaches to MoE models, due to the increased ratio of data to computation. In this paper, we introduce APTMoE, which employs affinityaware pipeline parallelism for fine-tuning MoE models on bandwidth-constrained GPU nodes. We propose an affinity-aware offloading technique that enhances pipeline parallelism for both computational efficiency and model size, and it benefits from a hierarchical loading strategy and a demand-priority scheduling strategy. To improve the computation efficiency and reduce the data movement volume, the hierarchical loading strategy designs three loading phases and efficiently allocates computation across GPUs and CPUs during these phases, leveraging different levels of expert popularity and computation affinity. With the aim of alleviating the mutual interference among the three loading phases and maximizing the bandwidth utilization, the demand-priority scheduling strategy proactively and dynamically coordinates the loading execution order. Experiments demonstrate that APTMoE outperforms existing methods in most cases. Particularly, APTMoE successfully fine-tunes a 61.2B MoE model on 4 Nvidia A800 GPUs(40GB) and achieves up to $33 \%$ throughput improvement compared to the SOTA method.

How to resolve the metabolic dark matter of microorganisms has long been a challenging problem in discovering active molecules. Diverse omics tools have been developed to guide the discovery and characterization of various microbial metabolites, which make it gradually possible to predict the overall metabolites for individual strains. The combinations of multi-omic analysis tools effectively compensates for the shortcomings of current studies that focus only on single omics or a broad class of metabolites. In this review, we systematically update, categorize and sort out different analysis tools for microbial metabolites prediction in the last five years to appeal for the multi-omic combination on the understanding of the metabolic nature of microbes. First, we provide the general survey on different updated prediction databases, webservers, or software that based on genomics, transcriptomics, proteomics, and metabolomics, respectively. Then, we discuss the essentiality on the integration of multi-omics data to predict metabolites of different microbial strains and communities, as well as stressing the combination of other techniques, such as systems biology methods and data-driven algorithms. Finally, we identify key challenges and trends in developing multi-omic analysis tools for more comprehensive prediction on diverse microbial metabolites that contribute to human health and disease treatment.

The mapping of tensor computation is a complex and important process for spatial accelerators. Today's mapping works depend on hand-tuned kernel libraries or search-based heuristics from human experts. The former is time-intensive while the latter easily leads to sub-optimal performance. In this paper, we propose TensorMap, a deep reinforcement learning (RL)-based mapping framework for tensor computations on spatial accelerators. We propose a sequential generation mode for mapping optimization and construct a coarse-grained action space to reduce the complexity of the mapping search space. An efficient policy network is devised to optimize mapping primitives in the RL-based search. We then propose a stop signal that is sampled fromBernoullidistribution to facilitate multi-level loop unrolling for spatial accelerators. Finally, a genetic algorithm is employed to further refine the optimized mappings. In the experiments, we demonstrate TensorMap's ability for different spatial accelerators with various tensor computations. On TPU, TensorMap provides 2.6$\times$, 2.7$\times$, and 2.4$\times$better energy-delay product (EDP) on average compared with FlexTensor, Ansor, and AMOS respectively. On Eyeriss, TensorMap provides 2.1$\times$, 1.8$\times$, and 1.7$\times$better EDP on average compared with FlexTensor, Ansor, and AMOS respectively.

Recommendation systems are essential to the operation of the majority of internet services, with Deep Learning Recommendation Models (DLRMs) serving as a crucial component. However, due to distinct computation, data access, and memory usage characteristics of recommendation models, the trainning of DLRMs may suffer from low resource utilization on prevalent heterogeneous CPU-GPU hardware platforms. Furthermore, as the majority of high-performance computing systems presently depend on multi-GPU computing nodes, the challenge of addressing low resource utilization becomes even more pronounced. Existing concurrent training solutions cannot be straightforwardly applied to DLRM due to various factors, such as insufficient fine-grained memory management and the lack of collaborative CPU-GPU scheduling. In this paper, we introduce RMixer, a scheduling framework that addresses these challenges by providing an efficient job management and scheduling mechanism for DLRM training jobs on heterogeneous CPU-GPU platforms. To facilitate training co-location, we first estimate the peak memory consumption of each job. Additionally, we track and collect resource utilization for DLRM training jobs. Based on the information of computational patterns, a batched job dispatcher with dynamic resource-complementary scheduling policy is proposed to co-locate DLRM training jobs on CPU-GPU platform. Scheduling strategies for both intra-GPU and inter-GPU scenarios were meticulously devised, with a focus on thoroughly examining individual GPU resource utilization and achieving a balanced state across multiple GPUs. Experimental results demonstrate that our implementation achieved up to 5.3× and 7.5× higher throughput on single GPU and 4 GPU respectively for training jobs involving various recommendation models.

Bladder cancer (BCa) is one of the most common malignancies in the world, which can be categorized into muscleinvasive (MIBC) and non-muscle-invasive (NMIBC). These two types of BCa must be treated differently, and thus it is essential to correctly distinguish MIBC and NMIBC patients preoperatively for adopting different treatment methods accordingly. Currently, the two types can be distinguished through MRI images by radiologists, but manual inspection is time and labor-consuming. Existing machine learning based methods attempt to free radiologists from manual inspection. However, they fail to take full advantage of image features and always require extra laborious refined manual labeling in addition to the classification labels. In this study, we propose a Tumor Staging and Localization Network (TSLNet) to perform preoperative non-invasive assessment of muscle invasion of BCa, which can automatically distinguish MIBC patients from NMIBC patients based on MRI T2-weighted images of BCa. The model adopts the weakly supervised learning method. Specifically, self-produced guidance is used as pixellevel segmentation pseudo labels for auxiliary supervision to extract basic features, and location-recognition based fine-grained image classification technology and inexact consistency labels are used for auxiliary supervision to extract fine-grained features. Moreover, the model can visualize the critical regions of the lesions, which can provide practical reference and a basis for clinicians’ clinical diagnosis. Experimental results show that the model achieves high AUC, accuracy, specificity, sensitivity, and F1-score, which is comparable to experienced clinicians.

Many applications such as biomedical analysis and scientific data analysis involve analyzing volumetric data. This spawns huge demand for 3D CNN. Although accelerators such as GPU may provide higher throughput on deep learning applications, they may not be available in all scenarios. CPU, especially many-core CPU, remains an attractive choice for deep learning in many scenarios. In this paper, we propose a inference solution that targets on the emerging ARM many-core CPU platform. A hierarchical partition approach is claimed to accelerate 3D-CNN inference by exploiting characteristics of memory and cache on ARM many-core CPU.

The development of deep learning recommendation models (DLRM) and recommendation systems has significantly improved the precision of information matching. Due to distinct computation, data access, and memory usage characteristics of recommendation models, they may suffer from low resource utilization on prevalent heterogeneous CPU-GPU hardware platforms. Existing concurrent training solutions cannot be directly applied to DLRM due to various factors, such as insufficient fine-grained memory management and the lack of collaborative CPU-GPU scheduling. In this paper, we introduce MixRec, a scheduling framework that addresses these challenges by pro-viding an efficient job management and scheduling mechanism for DLRM training jobs on heterogeneous CPU-GPU platforms. To facilitate training co-location, we first estimate the peak memory consumption of each job. Additionally, we track and collect resource utilization for DLRM training jobs. Based on the information of resource usage, a batched job dispatcher with dynamic resource-complementary scheduling policy is proposed to co-locate DLRM training jobs on CPU-GPU platform. Experimental results demonstrate that our implementation achieved up to 4.42× higher throughput and 3.97× higher resource utilization for training jobs involving various recommendation models.

Nowadays, Graphics Processing Units (GPUs) dominate in a wide spectrum of computing realms and multi-task is increasingly applied in various complicated applications. To gain higher performance, multi-task programs require cumbersome programming efforts to take advantage of inter-kernel concurrency at source-code level. Although there exist works automatically scheduling kernels to enable inter-kernel concurrency, they all inevitably introduce new programming frameworks and some even bring significant performance downgrade compared to the expertise-based optimizations. To address this issue, we propose KeSCo, a compiler-based scheduler to expose kernel level concurrency in multi-task programs with trivial code modification. In compilation, KeSCo applies a strategy to schedule kernels in task queues, accounting for both load balance and synchronization cost. Also, KeSCo utilizes a customized algorithm designed for computational flow to remove redundant synchronizations. The design is further extended to support multi-process scenario, where multiple GPU processes are sharing a single context. Evaluations on representative benchmarks show that the proposed approach gains a 1.28× average speedup for multi-task scenario (1.22× for multi-process). Even with lessened programming efforts, our proposed design outperforms two state-of-the-arts GrSched and Taskflow by 1.31× and 1.16× on average, respectively.

Dragonfly is a highly scalable, low-diameter, and cost-efficient network topology, which has been adopted in new exascale High Performance Computing (HPC) systems. However, Dragonfly topology suffers from the limited direct links between groups. The reconfigurable network can solve this problem by reconfiguring topology to adjust the number of direct links between groups. While the performance improvement of a single job on reconfigurable HPC network has been evaluated in previous works, the performance of HPC workloads has not been studied because of the lack of an appropriate resource allocation policy.

Weakly Supervised Object Localization (WSOL) aims to localize objects using only image-level labels while ensuring competitive classification performance. However, previous efforts have prioritized localization over classification accuracy in discriminative features, in which low-level information is neglected. We argue that low-level image representations, such as edges, color, texture, and motions are crucial for accurate detection. That is, using such information further achieves more refined localization, which can be used to promote classification accuracy. In this paper, we propose a unified framework that simultaneously improves localization and classification accuracy, termed as LocLoc (Low-level Cues and Local-area Guides). It leverages low-level image cues to explore global and local representations for accurate localization and classification. Specifically, we introduce a GrabCut-Enhanced Generator (GEG) to learn global semantic representations for localization based on graph cuts to enhance low-level information based on long-range dependencies captured by the transformer. We further design a Local Feature Digging Module (LFDM) that utilizes low-level cues to guide the learning route of local feature representations for accurate classification. Extensive experiments demonstrate the effectiveness of LocLoc with 84.4%(↑5.2%) Top-1 Loc., 85.8% Top-1 Cls. on CUB-200-2011 and 57.6% (↑1.5%) Top-1 Loc., 78.6% Top-1Cls. on ILSVRC 2012, indicating that our method achieves competitive performance with a large margin compared to previous approaches. Code and models are available at https://github.com/Cliffia123/LocLoc.

Crystal Structure Prediction (CSP) is crucial in various scientific disciplines. While CSP can be addressed by employing currently-prevailing generative models (**e.g.** diffusion models), this task encounters unique challenges owing to the symmetric geometry of crystal structures---the invariance of translation, rotation, and periodicity. To incorporate the above symmetries, this paper proposes DiffCSP, a novel diffusion model to learn the structure distribution from stable crystals. To be specific, DiffCSP jointly generates the lattice and atom coordinates for each crystal by employing a periodic-E(3)-equivariant denoising model, to better model the crystal geometry. Notably, different from related equivariant generative approaches, DiffCSP leverages fractional coordinates other than Cartesian coordinates to represent crystals, remarkably promoting the diffusion and the generation process of atom positions. Extensive experiments verify that our DiffCSP remarkably outperforms existing CSP methods, with a much lower computation cost in contrast to DFT-based methods. Moreover, the superiority of DiffCSP is still observed when it is extended for ab initio crystal generation.

Recent advances in spatial transcriptomics have enabled measurements of gene expression at cell/spot resolution meanwhile retaining both the spatial information and the histology images of the tissues. Accurately identifying the spatial domains of spots is a vital step for various downstream tasks in spatial transcriptomics analysis. To remove noises in gene expression, several methods have been developed to combine histopathological images for data analysis of spatial transcriptomics. However, these methods either use the image only for the spatial relations for spots, or individually learn the embeddings of the gene expression and image without fully coupling the information. Here, we propose a novel method ConGI to accurately exploit spatial domains by adapting gene expression with histopathological images through contrastive learning. Specifically, we designed three contrastive loss functions within and between two modalities (the gene expression and image data) to learn the common representations. The learned representations are then used to cluster the spatial domains on both tumor and normal spatial transcriptomics datasets. ConGI was shown to outperform existing methods for the spatial domain identification. In addition, the learned representations have also been shown powerful for various downstream tasks, including trajectory inference, clustering, and visualization.

SUMMARY: VSTH is a user-friendly web server with the complete workflow for virtual screening. By self-customized visualization software, users can interactively prepare protein files, set docking sites as well as view binding conformers in a target protein in a few clicks. We provide serval purchasable ligand libraries for selection. And, we integrate six open-source docking programs as computing engine, or as conformational sampling tools for DLIGAND2. Users can select various docking methods simultaneously and personalize computing parameters. After docking processing, user can filter docking conformations by ranked scores, or cluster-based molecular similarity to find highly populated clusters of low-energy conformations. AVAILABILITY AND IMPLEMENTATION: The VSTH web server is free and open to all users at https://matgen.nscc-gz.cn/VirtualScreening.html. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Identifying the B-cell epitopes is an essential step for guiding rational vaccine development and immunotherapies. Since experimental approaches are expensive and time-consuming, many computational methods have been designed to assist B-cell epitope prediction. However, existing sequence-based methods have limited performance since they only use contextual features of the sequential neighbors while neglecting structural information. RESULTS: Based on the recent breakthrough of AlphaFold2 in protein structure prediction, we propose GraphBepi, a novel graph-based model for accurate B-cell epitope prediction. For one protein, the predicted structure from AlphaFold2 is used to construct the protein graph, where the nodes/residues are encoded by ESM-2 learning representations. The graph is input into the edge-enhanced deep graph neural network (EGNN) to capture the spatial information in the predicted 3D structures. In parallel, a bidirectional long short-term memory neural networks (BiLSTM) are employed to capture long-range dependencies in the sequence. The learned low-dimensional representations by EGNN and BiLSTM are then combined into a multilayer perceptron for predicting B-cell epitopes. Through comprehensive tests on the curated epitope dataset, GraphBepi was shown to outperform the state-of-the-art methods by more than 5.5% and 44.0% in terms of AUC and AUPR, respectively. A web server is freely available at http://bio-web1.nscc-gz.cn/app/graphbepi. AVAILABILITY AND IMPLEMENTATION: The datasets, pre-computed features, source codes, and the trained model are available at https://github.com/biomed-AI/GraphBepi.

Vision Transformer (ViT) has shown great potential for various visual tasks due to its ability to model long-range dependency. However, ViT requires a large amount of computing resource to compute the global self-attention. In this work, we propose a ladder self-attention block with multiple branches and a progressive shift mechanism to develop a light-weight transformer backbone that requires less computing resources (e.g., a relatively small number of parameters and FLOPs), termed Progressive Shift Ladder Transformer (PSLT). First, the ladder self-attention block reduces the computational cost by modelling local self-attention in each branch. In the meanwhile, the progressive shift mechanism is proposed to enlarge the receptive field in the ladder self-attention block by modelling diverse local self-attention for each branch and interacting among these branches. Second, the input feature of the ladder self-attention block is split equally along the channel dimension for each branch, which considerably reduces the computational cost in the ladder self-attention block (with nearly [Formula: see text] the amount of parameters and FLOPs), and the outputs of these branches are then collaborated by a pixel-adaptive fusion. Therefore, the ladder self-attention block with a relatively small number of parameters and FLOPs is capable of modelling long-range interactions. Based on the ladder self-attention block, PSLT performs well on several vision tasks, including image classification, objection detection and person re-identification. On the ImageNet-1 k dataset, PSLT achieves a top-1 accuracy of 79.9% with 9.2 M parameters and 1.9 G FLOPs, which is comparable to several existing models with more than 20 M parameters and 4 G FLOPs. Code is available at https://isee-ai.cn/wugaojie/PSLT.html.

Transformer models have emerged as a leading approach in the field of natural language processing (NLP) and are increasingly being deployed in production environments. Graphic processing units (GPUs) have become a popular choice for the transformer deployment and often rely on the batch processing technique to ensure high hardware performance. Nonetheless, the current practice for transformer inference encounters computational and memory redundancy due to the heavy-tailed distribution of sequence lengths in NLP scenarios, resulting in low practical performance. In this article, we propose a unified solution for improving both computation and memory efficiency of the real-world transformer inference on GPUs. The solution eliminates the redundant computation and memory footprint across a transformer model. At first, a GPU-oriented computation approach is proposed to process the self-attention module in a fine-grained manner, eliminating its redundant computation. Next, the multi-layer perceptron module continues to use the word-accumulation approach to eliminate its redundant computation. Then, to better unify the fine-grained approach and the word-accumulation approach, it organizes the data layout of the self-attention module in block granularity. Since aforementioned approaches make the required memory size largely reduce and constantly fluctuate, we propose the chunk-based approach to enable a better balance between memory footprint and allocation/free efficiency. Our experimental results show that our unified solution achieves a decrease of average latency by 28% on the entire transformer model, 63.8% on the self-attention module, and reduces memory footprint of intermediate results by 7.8×, compared with prevailing frameworks.

The tremendous success of convolutional neural network (CNN) has made it ubiquitous in many fields of human endeavor. Many applications such as biomedical analysis and scientific data analysis involve analyzing volumetric data. This spawns huge demand for 3D-CNN. Although accelerators such as GPU may provide higher throughput on deep learning applications, they may not be available in all scenarios. CPU, especially many-core CPU with non-uniform memory access (NUMA) architecture, remains an attractive choice for deep learning inference in many scenarios. In this article, we propose a distributed inference solution for 3D-CNN that targets on the emerging ARM many-core CPU platform. A hierarchical partition approach is claimed to accelerate 3D-CNN inference by exploiting characteristics of memory and cache on ARM many-core CPU. Based on the hierarchical model partition approach, other optimization techniques such as NUMA-aware thread scheduling and optimization of 3D-img2row convolution are designed to exploit the potential of ARM many-core CPU for 3D-CNN. We evaluate our proposed inference solution with several classic 3D-CNNs: C3D, 3D-resnet34, 3D-resnet50, 3D-vgg11, and P3D. Our experimental results show that our solution can boost the performance of the 3D-CNN inference, and achieve much better scalability, with a negligible fluctuation in accuracy. When employing our 3D-CNN inference solution on ACL libraries, it can outperform naive ACL implementations by 11× to 50× on ARM many-core processor. When employing our 3D-CNN inference solution on NCNN libraries, it can outperform the naive NCNN implementations by 5.2× to 14.2× on ARM many-core processor.

Malware detection approaches have been extensively studied for traditional software systems. However, the development of blockchain technology has promoted the birth of a new type of software system–decentralized applications. Composed of smart contracts, a type of application that implements the Ponzi scheme logic (called smart Ponzi schemes) has caused irreversible loss and hindered the development of blockchain technology. These smart contracts generally had a short life but involved a large amount of money. Whereas identification of these Ponzi schemes before causing financial loss has been significantly important, existing methods suffer from three main deficiencies, i.e., the insufficient dataset, the reliance on the transaction records, and the low accuracy. In this study, we first build a larger dataset. Then, a large number of features from multiple views, including bytecode, semantic, and developers, are extracted. These features are independent of the transaction records. Furthermore, we leveraged machine learning methods to build our identification model, i.e., Mul ti-view Cas cade Ensemble model (MulCas). The experiment results show that MulCas can achieve higher performance and robustness in the scope of our dataset. Most importantly, the proposed method can identify smart Ponzi scheme at the creation time.

The past several years have witnessed tremendous success of transformer models in natural language processing (NLP), and their current landscape is increasingly diverse. Although GPU gradually becomes the dominating workhorse and de facto standard for deep learning, there are still many scenarios where using CPU remains a prevalent choice.Recently, ARM many-core processor starts emigrating to cloud computing and high-performance computing, which is promising to deploy transformer inference. In this paper, we identify several performance bottlenecks of existing inference runtime on many-core CPU including low-core usage, isolated thread configuration, inappropriate implementation of general matrix multiply (GEMM), and redundant computations for variable-length inputs. To tackle these problems, full-stack optimizations are conducted for these challenges from service level to operator level. We explore multi-instance parallelization at the service level to improve CPU core usage. To improve parallel efficiency of the inference runtime, we design NUMA-aware thread scheduling and a look-up table for optimal parallel configurations. The GEMM implementation is tailored for some critical modules to exploit the characteristics of transformer workload. To eliminate redundant computations, a novel storage format is designed and implemented to pack sparse data and a load balancing strategy is proposed for tasks with different sparsity. Experiments show that our implementation can outperform existing solutions by 1.1x to 6x with fixed-length inputs. For variable-length inputs, it achieves 1.9x to 8x speedups on different ARM many-core processors.

As the de facto high-throughput accelerators, graphics processing units (G PU s) are now used in a wide spec-trum of fields, including artificial intelligence, high performance computing and finance. While with excessive computing and memory resources, G PU s are facing significant challenges to reach high utilization by a monolithic task. Multiple tasks are thus concurrently running to share the GPUs, but they may adversely affect each other, causing performance degradation. As a result, it is extremely critical to manage resources in a reasonable way to strike a balance between utilization and performance. Targeting the issue, this paper proposes an effective resource management design via hybrid task scheduling. Our design continuously tracks the G PU executions and collects the usage statistics, which are then used to direct the task selection and dispatch, including the type, starting time and kernel dimensions. A prototype is developed on off-the-shelf GPUs by moderately refactoring the CUDA source codes. Experimental results show that the design can achieve up to 1.96x performance improvement (1.51x on average), meanwhile effectively boosting resource utilization.

Arithmetic operators are now used in a wide spectrum of domains, including artificial intelligence, data analytics and scientific computing. Meanwhile, specialized hardware components to enable low-precision computing are increasingly deployed in GPUs and accelerators. Whereas promising to boost performance, accelerating the operators on the hardware necessitates manually tuning the mixed-precision knobs to balance the performance and accuracy, which can be extremely challenging in real practices.

This paper proposes a locality-aware Oblivious RAM (ORAM) primitive, named Green ORAM, which exploits spatial locality of data in the physical memory for reducing ORAM overheads. The Green ORAM is novel consisting of three policies. The first is row-guided label allocation used for mapping spatial locality onto ORAM tree to reduce the number of memory commands. The second is segment-based path replacement able to improve the data locality within the path in the ORAM tree in order to remove the redundant memory accesses. The third is multi-path write-back able to improve the data locality between different paths in order to obtain theoretical best stash hit rate. Notably, the Green ORAM still maintains the security as we analyzed. Experimental results show that Green ORAM achieves a 28.72% access latency reduction, and a 19.06% memory energy consumption reduction on average, compared with the state-of-the-art String ORAM.

Transformer has experienced tremendous success and revolutionized the field of natural language processing (NLP). While GPU has become the de facto standard for deep learning computation in many cases, there are still many scenarios where using CPU for deep learning remains a prevalent choice. In particular, ARM many-core processor is emerging as a competitive candidate for HPC systems, which is promising to deploy Transformer inference.

Transformer-based models achieve superior accuracy in the field of natural language processing (NLP) and start to be widely deployed in production. As a popular deployment device, graphic processing units (GPUs) basically adopt the batch processing technique for inferring transformer-based models and achieving high hardware performance. However, as the input sequence lengths of NLP tasks are generally variable and in a heavy-tailed distribution, the batch processing will bring large amounts of redundant computation and hurt the practical efficiency.

The Message Passing Interface (MPI) is the most prominent and dominant programming model for scientific computing in super-computing systems today. Although many general and efficient algorithms have been proposed for MPI collective operations, there is still room for topology-aware optimization. Dragonfly is a high-scalability, low-diameter, and cost-efficient network topology adopted in more and more supercomputing networks. However, Dragonfly topology limits the performance of some MPI collective operations. In this paper, our analysis shows that the bottlenecks of collective algorithms in Dragonfly topology are intra-job interference, inter-job interference, and topology mismatch. We propose 5 different optimizations, i.e., Pseudo-random Pairwise, Tree-based Shuffle, Reversed Recursive Doubling, Reordered Bruck, and Matched Rabenseifner, for MPI collective operations including All-Gather, All-to-All, All-Reduce, and Reduce-Scatter. We evaluate each optimization through CODES network simulation framework with minimal, non-minimal, and adaptive routing. The simulation results demonstrate that the performance of All-to-All, All-Gather, All-Reduce, and Reduce-Scatter can be improved by 4.7X, 3.4X, 12.7%, and 4.1X, respectively, for 32768-node jobs with adaptive routing.

Code size is an increasing concern on resource constrained systems, ranging from embedded devices to cloud servers. To address the issue, lowering memory occupancy has become a priority in developing and deploying applications, and accordingly compiler-based optimizations have been proposed to reduce program footprint. However, prior arts are generally dealing with source codes or intermediate representations, and thus are very limited in scope in real scenarios where only binary files are commonly provided. To fill the gap, this paper presents a novel code-size optimization RollBin to reroll loops at binary level. RollBin first locates the unrolled loops in binary files, and then probes to decide the unrolling factor by identifying regular memory address patterns. To reconstruct the iterations, we propose a customized data dependency analysis that tackles the challenges brought by shuffled instructions and loop-carry dependencies. Next, the recognized iterations are rolled up through instruction removal and update, which are generally reverting the normal unrolling procedure. The evaluations on standard SPEC2006/2017 and MiBench demonstrate that RollBin effectively shrinks code size by 1.7% and 2.2% on average (up to 7.8%), which respectively outperforms the state-of-the-arts by 31% and 38%. In addition, the use cases of representative realistic applications manifest that RollBin can be applicable in practices.

Self-supervised pre-training techniques have achieved remarkable progress in Document AI. Most multimodal pre-trained models use a masked language modeling objective to learn bidirectional representations on the text modality, but they differ in pre-training objectives for the image modality. This discrepancy adds difficulty to multimodal representation learning. In this paper, we propose LayoutLMv3 to pre-train multimodal Transformers for Document AI with unified text and image masking. Additionally, LayoutLMv3 is pre-trained with a word-patch alignment objective to learn cross-modal alignment by predicting whether the corresponding image patch of a text word is masked. The simple unified architecture and training objectives make LayoutLMv3 a general-purpose pre-trained model for both text-centric and image-centric Document AI tasks. Experimental results show that LayoutLMv3 achieves state-of-the-art performance not only in text-centric tasks, including form understanding, receipt understanding, and document visual question answering, but also in image-centric tasks such as document image classification and document layout analysis. The code and models are publicly available at https://aka.ms/layoutlmv3.

Single-cell RNA sequencing (scRNA-seq) techniques provide high-resolution data on cellular heterogeneity in diverse tissues, and a critical step for the data analysis is cell type identification. Traditional methods usually cluster the cells and manually identify cell clusters through marker genes, which is time-consuming and subjective. With the launch of several large-scale single-cell projects, millions of sequenced cells have been annotated and it is promising to transfer labels from the annotated datasets to newly generated datasets. One powerful way for the transferring is to learn cell relations through the graph neural network (GNN), but traditional GNNs are difficult to process millions of cells due to the expensive costs of the message-passing procedure at each training epoch. Here, we have developed a robust and scalable GNN-based method for accurate single-cell classification (GraphCS), where the graph is constructed to connect similar cells within and between labelled and unlabeled scRNA-seq datasets for propagation of shared information. To overcome the slow information propagation of GNN at each training epoch, the diffused information is pre-calculated via the approximate Generalized PageRank algorithm, enabling sublinear complexity over cell numbers. Compared with existing methods, GraphCS demonstrates better performance on simulated, cross-platform, cross-species and cross-omics scRNA-seq datasets. More importantly, our model provides a high speed and scalability on large datasets, and can achieve superior performance for 1 million cells within 50 min.

The rapid development of spatial transcriptomics allows the measurement of RNA abundance at a high spatial resolution, making it possible to simultaneously profile gene expression, spatial locations of cells or spots, and the corresponding hematoxylin and eosin-stained histology images. It turns promising to predict gene expression from histology images that are relatively easy and cheap to obtain. For this purpose, several methods are devised, but they have not fully captured the internal relations of the 2D vision features or spatial dependency between spots. Here, we developed Hist2ST, a deep learning-based model to predict RNA-seq expression from histology images. Around each sequenced spot, the corresponding histology image is cropped into an image patch and fed into a convolutional module to extract 2D vision features. Meanwhile, the spatial relations with the whole image and neighbored patches are captured through Transformer and graph neural network modules, respectively. These learned features are then used to predict the gene expression by following the zero-inflated negative binomial distribution. To alleviate the impact by the small spatial transcriptomics data, a self-distillation mechanism is employed for efficient learning of the model. By comprehensive tests on cancer and normal datasets, Hist2ST was shown to outperform existing methods in terms of both gene expression prediction and spatial region identification. Further pathway analyses indicated that our model could reserve biological information. Thus, Hist2ST enables generating spatial transcriptomics data from histology images for elucidating molecular signatures of tissues.

Clustering analysis is widely used in single-cell ribonucleic acid (RNA)-sequencing (scRNA-seq) data to discover cell heterogeneity and cell states. While many clustering methods have been developed for scRNA-seq analysis, most of these methods require to provide the number of clusters. However, it is not easy to know the exact number of cell types in advance, and experienced determination is not always reliable. Here, we have developed ADClust, an automatic deep embedding clustering method for scRNA-seq data, which can accurately cluster cells without requiring a predefined number of clusters. Specifically, ADClust first obtains low-dimensional representation through pre-trained autoencoder and uses the representations to cluster cells into initial micro-clusters. The clusters are then compared in between by a statistical test, and similar micro-clusters are merged into larger clusters. According to the clustering, cell representations are updated so that each cell will be pulled toward centers of its assigned cluster and similar clusters, while cells are separated to keep distances between clusters. This is accomplished through jointly optimizing the carefully designed clustering and autoencoder loss functions. This merging process continues until convergence. ADClust was tested on 11 real scRNA-seq datasets and was shown to outperform existing methods in terms of both clustering performance and the accuracy on the number of the determined clusters. More importantly, our model provides high speed and scalability for large datasets.

Abstract Cloud storage is a fundamental component of the cloud computing system, which significantly affects the overall performance and quality of service of the cloud. Cloud storage servers face the challenge of imbalanced workloads. According to our observations on the time series generated by cloud storage, we found that the imbalance workloads will dramatically increase the tail latency of data access in the multi-tenant scenario. The intuitive solution is to periodicity detect the imbalance storage nodes and re-balance the loads. However, there are four challenges to accurately detect load of storage in the cloud with multiple tenants since the load may change frequently in cloud. This paper proposes PrecisePeriod, a precise periodicity detection algorithm customized for multi-tenant cloud storage. It removes outliers through data preprocessing, employs the discrete wavelet transform to remove high-frequency noise while keeping frequency domain information, computes the candidate periodicity queue using the autocorrelation function, and determines precise period through periodicity verification. Then, we design a cloud storage load balancing scheduling strategy based on PrecisePeriod, and the evaluation shows that the PrecisePeriod scheduling significantly reduces tail latency while only bringing $$1-2\%$$ 1-2% overhead.

Convolutional neural networks (CNNS) enable machines to view the world as humans and become increasing prevalent for Internet of Things (IoT) applications. Instead of streaming the raw data to the cloud and executing CNN inference remotely, it would be very attractive to use local IoT devices to process as it enables IoT applications with independent decision-making ability. Since a single IoT device can hardly match the requirements of the CNN inference, especially for time-sensitive and high-accuracy tasks, the distributedin-situCNN inference becomes a potential solution. However, because of the inherently tightly coupled structure of existing CNN models, it is difficult to distribute the inference efficiently. In this article, we enhance the distributedin-situCNN inference in the IoT. We fundamentally reduce the communication overhead of distributed CNN inference by designing new loosely coupled structure (LCS). Experimental results demonstrate that LCS achieves the leading performance compared with other popular structures. Next, based on the LCS, we customize the partitioning method to reduce the synchronization points and design the decentralized asynchronous method to optimize communication in each synchronization point. To evaluate the effectiveness, we build a prototype system. When the number of IoT devices increases from 1 to 4, our system accelerates by up to$3.85\times $and reduces the memory footprint in each device by 70% with achieving a competitive accuracy and significantly outperforming other approaches.

MOTIVATION: The interactions of proteins with DNA, RNA, peptide, and carbohydrate play key roles in various biological processes. The studies of uncharacterized protein-molecules interactions could be aided by accurate predictions of residues that bind with partner molecules. However, the existing methods for predicting binding residues on proteins remain of relatively low accuracies due to the limited number of complex structures in databases. As different types of molecules partially share chemical mechanisms, the predictions for each molecular type should benefit from the binding information with other molecule types. RESULTS: In this study, we employed a multiple task deep learning strategy to develop a new sequence-based method for simultaneously predicting binding residues/sites with multiple important molecule types named MTDsite. By combining four training sets for DNA, RNA, peptide, and carbohydrate-binding proteins, our method yielded accurate and robust predictions with AUC values of 0.852, 0836, 0.758, and 0.776 on their respective independent test sets, which are 0.52 to 6.6% better than other state-of-the-art methods. To my best knowledge, this is the first method using multi-task framework to predict multiple molecular binding sites simultaneously.

Phase Change Memory (PCM) can directly connect persistent memory to main memory bus, while it achieves high read throughput and low standby power, the critical concerns are its poor write performance and limited durability. A naturally in-spired design is the hybrid memory architecture that fuses DRAM and PCM, so as to exploit the positive aspects of both types of memory. Unfortunately, existing solutions are seriously challenged by the limited main memory size, which is the primary bottleneck of in-memory computing. In this paper, we introduce a novel Content Aware Hybrid DRAM-PCM memory system framework—CAHRAM, which exploits deduplication to improve line sharing with high memory efficiency. It reduces write traffic to hybrid memory by removing unnecessary duplicate line writes, thereby further enhancing the write endurance of PCM. And it also substantially extends available free memory space by coalescing redundant lines in hybrid memory. We also design a reference-based page migration technique to minimize the access overheads caused by the performance gap between DRAM and PCM. Compared with the state-of-the-art in a hybrid memory simulator, our experiment results show that CAHRAM can achieve the highest I/O performance and the longest PCM lifetime with the competitive efficiencies in space and energy.

The postprocessing method of ensemble forecasts is usually used to find a more precise estimate of future precipitation, because dynamic meteorology models have limitations in fitting fine-grained atmospheric processes and precipitation is driven more often by smaller-scale processes, while ensemble forecasts can hit this precipitation at times. However, the pattern of these hits cannot be easily summarized. The existing objective postprocessing methods tend to extend the rain area or false alarm the precipitation intensity categories. In this work, we introduce a multi-layer structure to simultaneously reduce the bias in forecast ensembles output by meteorology models and merge them to a quality deterministic (single-valued) forecast using cross-grid information, which differs quite dramatically from the previous statistical postprocessing method. The multi-layer network is designed to model the spatial distribution of future precipitation of different intensity categories(IC-MLNet). We provide a comparison of IC-MLNet to simple average as well as another two state-of-the-art ensemble quantitative precipitation forecasts (QPFs) postprocessing approaches over both single-model and multi-model ensemble forecasts datasets from TIGGE. The experimental results indicate that our model achieves superior performance over the compared baselines in precipitation amount prediction as well as precipitation intensities categories prediction.

It is a common strategy to predict oncogenes by differential analysis between somatic mutations and background mutations. Most previous methods only utilize mutations in the cancer population to model its background mutation, which have an obvious bias. A recent method, DiffMut, improves this issue by conducting differential mutational analysis with both mutations in the cancer population and the natural population. However, it assumes the impacts of all mutations are equal, neglecting their functional difference. Thus, we developed a method, DGAT-onco that integrated the functional impacts of mutations to the differential mutational analysis framework of DiffMut. We performed DGAT-onco analysis with 33 cancer types from the Cancer Genome Atlas (TCGA) dataset. Its reliability was further evaluated on an independent test set including 22 cancers from other sources (TS22). Using oncogenes from the Cancer Gene Census (CGC) as the gold standard, our method achieves higher classification performance in oncogene discovery than five alternative methods (i.e., DiffMut, WITER, OncodriveCLUSTL, OncodriveFML, and MutSigCV) with an average AUPRC of 0.197 and 0.187 in TCGA and TS22 respectively. The source code and supplementary materials of DGAT-onco are available at https://github.com/zhanghaoyang0/DGAT-onco.

Vision-Dialog Navigation (VDN) requires an agent to ask questions and navigate following the human responses to find target objects. Conventional approaches are only allowed to ask questions at predefined locations, which are built upon expensive dialogue annotations, and inconvenience the real-word human-robot communication and cooperation. In this paper, we propose a Self-Motivated Communication Agent (SCoA) that learns whether and what to communicate with human adaptively to acquire instructive information for realizing dialogue annotation-free navigation and enhancing the transferability in real-world unseen environment. Specifically, we introduce a whether-to-ask (WeTA) policy, together with uncertainty of which action to choose, to indicate whether the agent should ask a question. Then, a what-to-ask (WaTA) policy is proposed, in which, along with the oracle’s answers, the agent learns to score question candidates so as to pick up the most informative one for navigation, and meanwhile mimic oracle’s answering. Thus, the agent can navigate in a self-Q&A manner even in real-world environment where the human assistance is often unavailable. Through joint optimization of communication and navigation in a unified imitation learning and reinforcement learning framework, SCoA asks a question if necessary and obtains a hint for guiding the agent to move towards the target with less communication cost. Experiments on seen and unseen environments demonstrate that SCoA shows not only superior performance over existing baselines without dialog annotations, but also competing results compared with rich dialog annotations based counterparts.

Convolution Neural Network (CNN) has gained a great success in deep learning applications and been accelerated by dedicated convolutional algorithms. Winograd-based algorithm can greatly reduce the number of arithmetic operations required in convolution. However, our experiments show that existing implementations in deep learning libraries cannot achieve expected parallel performance on ARM manycore CPUs with last-level cache (LLC). Compared to multicore processor, ARM manycore CPUs have more cores, more NUMA nodes and the parallel performance is more easily restricted by memory bandwidth, cache contention, NUMA configuration and etc. In this paper, we propose an optimized implementation for single-precision Winograd-based algorithm on ARM manycore CPUs. Our algorithm adjusts the data layout according to the input shape and is optimized for the characteristics of ARM processor, thus reducing the matrix transformation overhead and achieving high arithmetic intensity. We redesign the parallel algorithm for Winograd-based convolution to achieve a more efficient implementation for manycore CPUs. The experimental results with 32 cores show that for modern ConvNets, our implementation achieves speedups ranging from 3 × to 5 × over the state-of-the-art Winograd-based convolution on ARM processor. Even conducted on a set of convolutional benchmarks executing on a 128-core system with 4 NUMA nodes, the results show that our implementation can also achieve better performance than existing implementations on ARM processor.

A creative image-and-text generative AI system mimics humans' extraordinary abilities to provide users with diverse and comprehensive caption suggestions, as well as rich image creations. In this work, we demonstrate such an AI creation system to produce both diverse captions and rich images. When users imagine an image and associate it with multiple captions, our system paints a rich image to reflect all captions faithfully. Likewise, when users upload an image, our system depicts it with multiple diverse captions. We propose a unified multi-modal framework to achieve this goal. Specifically, our framework jointly models image-and-text representations with a Transformer network, which supports rich image creation by accepting multiple captions as input. We consider the relations among input captions to encourage diversity in training and adopt a non-autoregressive decoding strategy to enable real-time inference. Based on these, our system supports both diverse captions and rich images generations. Our code is available online.

We study the joint learning of image-to-text and text-to-image generations, which are naturally bi-directional tasks. Typical existing works design two separate task-specific models for each task, which impose expensive design efforts. In this work, we propose a unified image-and-text generative framework based on a single multimodal model to jointly study the bi-directional tasks. We adopt Transformer as our unified architecture for its strong performance and task-agnostic design. Specifically, we formulate both tasks as sequence generation tasks, where we represent images and text as unified sequences of tokens, and the Transformer learns multimodal interactions to generate sequences. We further propose two-level granularity feature representations and sequence-level training to improve the Transformer-based unified framework. Experiments show that our approach significantly improves previous Transformer-based model X-LXMERT's FID from 37.0 to 29.9 (lower is better) for text-to-image generation, and improves CIDEr-D score from 100.9% to 122.6% for fine-tuned image-to-text generation on the MS-COCO dataset. Our code is available online.

As a large number of emerging graph applications spread across different domains, the need for processing massive concurrent graph jobs (CGJs) is increasing. However, existing graph processing systems designed for a single job cannot efficiently tackle multiple CGJs, where they suffer from interfering memory access patterns and inefficient property management. In this paper, we introduce Krill, a compiler and runtime system for processing concurrent graph jobs. We propose an SAP model, which decouples graph structure, algorithm, and property. In the compiler, we propose leveraging the property buffer to easily write and manage property data. In the runtime system, we propose a novel technique named graph kernel fusion to reduce memory accesses, which fuses all the jobs and processes them as a whole. Experimental results show our system significantly reduces the number of memory accesses for CGJs by more than 6x compared with the baseline, and achieves up to 6.76x speedup with 3.84x shorter response latency than GraphM, the state-of-the-art concurrent graph processing system.

It is promising to deploy CNN inference on local end-user devices for high-accuracy and time-sensitive applications. Model parallelism has the potential to provide high throughput and low latency in distributed CNN inference. However, it is non-trivial to use model parallelism as the original CNN model is inherently tightly-coupled structure. In this article, we propose DeCNN, a more effective inference approach that uses decoupled CNN structure to optimize model parallelism for distributed inference on end-user devices. DeCNN is novel consisting of three schemes. Scheme-1 is structure-level optimization. It exploits group convolution and channel shuffle to decouple the original CNN structure for model parallelism. Scheme-2 is partition-level optimization. It is based on channel group to partition the convolutional layers, and then leverages input-based method to partition the fully connected layers, further exposing high degree of parallelism. Scheme-3 is communication-level optimization. It uses inter-sample parallelism to hide communications for better performance and robustness, especially in the weak network connections. We use ImageNet classification task to evaluate the effectiveness of DeCNN on a distributed multi-ARM platform. Notably, when using the number of devices from 1 to 4, DeCNN can accelerate the inference of large-scale ResNet-50 by 3.21×, and reduce 65.3 percent memory footprint, with 1.29 percent accuracy improvement.

In this article, a parallel structured divide-and-conquer (PSDC) eigensolver is proposed for symmetric tridiagonal matrices based on ScaLAPACK and a parallel structured matrix multiplication algorithm, called PSMMA. Computing the eigenvectors via matrix-matrix multiplications is the most computationally expensive part of the divide-and-conquer algorithm, and one of the matrices involved in such multiplications is a rank-structured Cauchy-like matrix. By exploiting this particular property, PSMMA constructs the local matrices by using generators of Cauchy-like matrices without any communication, and further reduces the computation costs by using a structured low-rank approximation algorithm. Thus, both the communication and computation costs are reduced. Experimental results show that both PSMMA and PSDC are highly scalable and scale to 4096 processes at least. PSDC has better scalability than PHDC that was proposed in [16] and only scaled to 300 processes for the same matrices. Comparing with PDSTEDC in ScaLAPACK, PSDC is always faster and achieves 1.4x-1.6x speedup for some matrices with few deflations. PSDC is also comparable with ELPA, with PSDC being faster than ELPA when using few processes and a little slower when using many processes.

With the prevalence of web services, a large number of similar web services are provided by different providers. To select the optimal service among these service candidates, Quality of Service (QoS), representing the non-functional characteristics, plays an important role. To obtain the QoS values of web services, a number of web service QoS prediction methods have been proposed. Collaborative web service QoS prediction is one of the most popular approaches. Based on the historical QoS data, collaborative QoS prediction methods employ memory-based collaborative filtering (CF), model-based CF, or their hybrids to predict QoS values. However, these methods usually only consider the QoS information of similar users and services, neglecting the correlation between them. To enhance the prediction accuracy, we propose a novel method to predict QoS values based on factorization machine, which leverages not only QoS information of users and services but also the user and service neighbor’s information. To evaluate our approach, we conduct experiments on a large-scale real-world dataset with 1,974,675 web service invocations. The experiment results show that our approach achieves higher prediction accuracy than other QoS prediction methods.

The following topics are dealt with: learning (artificial intelligence); diseases; medical image processing; molecular biophysics; genetics; medical computing; feature extraction; cancer; genomics; proteins.

Recent advances in single-cell RNA sequencing (scRNA-seq) technologies provide a great opportunity to study gene expression at cellular resolution, and the scRNA-seq data has been routinely conducted to unfold cell heterogeneity and diversity. A critical step for the scRNA-seq analyses is to cluster the same type of cells, and many methods have been developed for cell clustering. However, existing clustering methods are limited to extract the representations from expression data of individual cells, while ignoring the high-order structural relations between cells. Here, we proposed a new method (GraphSCC) to cluster cells based on scRNA-seq data by accounting structural relations between cells through a graph convolutional network. The representation learned from the graph convolutional network, together with another representation output from a denoising autoencoder network, are optimized by a dual self-supervised module for better cell clustering. Extensive experiments indicate that GraphSCC model outperforms state-of-the-art methods in various evaluation metrics on both simulated and real datasets.

In recent years, blockchain technology has created a new cryptocurrency world and has attracted a lot of attention. It also is rampant with various scams. For example, phishing scams have grabbed a lot of money and has become an important threat to users' financial security in the blockchain ecosystem. To help deal with this issue, this paper proposes a systematic approach to detect phishing accounts based on blockchain transactions and take Ethereum as an example to verify its effectiveness. Specifically, we propose a graph-based cascade feature extraction method based on transaction records and a lightGBM-based Dual-sampling Ensemble algorithm to build the identification model. Extensive experiments show that the proposed algorithm can effectively identify phishing scams.

Constructing proper representations of molecules lies at the core of numerous tasks such as molecular property prediction and drug design. Graph neural networks, especially message passing neural network (MPNN) and its variants, have recently made remarkable achievements in molecular graph modeling. Albeit powerful, the one-sided focuses on atom (node) or bond (edge) information of existing MPNN methods lead to the insufficient representations of the attributed molecular graphs. Herein, we propose a Communicative Message Passing Neural Network (CMPNN) to improve the molecular embedding by strengthening the message interactions between nodes and edges through a communicative kernel. In addition, the message generation process is enriched by introducing a new message booster module. Extensive experiments demonstrated that the proposed model obtained superior performances against state-of-the-art baselines on six chemical property datasets. Further visualization also showed better representation capacity of our model.

Traditional distributed file systems (DFS) use centralized service to manage metadata. Many studies based on this centralized architecture enhanced metadata processing capability by scaling the metadata server cluster, which is however still difficult to keep up with the growing number of clients and the increasingly metadata-intensive applications. Some solutions abandoned the centralized metadata service and improved scalability by embedding a private metadata service in an HPC application, but these solutions are suitable for only some specific applications and the absence of global namespace makes data sharing and management difficult. This paper addresses the shortcomings of existing studies by optimizing the consistency model of client- side metadata cache for the HPC scenario using a novel partial consistency model. It provides the application with strong consistency guarantee for only its workspace, thus improving metadata scalability without adding hardware or sacrificing the versatility and manageability of DFSes. In addition, the paper proposes batch permission management to reduce path traversal overhead, thereby improving metadata processing efficiency. The result is a library (Pacon) that allows existing DFSes to achieve partial consistency for scalable and efficient metadata management. The paper also presents a comprehensive evaluation using intensive benchmarks and representative application. For example, in file creation, Pacon improves the performance of BeeGFS by more than 76.4 times, and outperforms the state-of-the-art metadata management solution (IndexFS) by more than 4.6 times.

As Ethereum's smart contracts have boomed, it has become an integral part of the blockchain ecosystem. Unfortunately, some malicious users also find the opportunity to use fraudulent means to profit. A new reported approach is to lure new users or other attackers into the contract in an attempt to make a profit by exposing seemingly obvious flaws in the contract. But in fact, the contract contains a hidden trap that ultimately benefits the creator of the contract. Such contracts are known as honeypot contracts in the blockchain ecosystem. Previous studies proposed two methods to identify such smart contracts by using symbolic execution and contract behaviors. However, these methods either make it difficult to discover new categories or fail to warn users before they lose money. To solve this problem, we propose a machine learning model to detect honeypot contracts based on N-gram features and LightGBM. Extensive experiments show that our proposed model performs well in different conditions.

In general, race condition can be resolved by introducing synchronisations or breaking data dependencies. Atomic operations and graph coloring are the two typical approaches to avoid race condition. Graph coloring algorithms have been generally considered winning algorithms in the literature due to their lock free implementations. In this paper, we present the GPU-accelerated algorithms of the unstructured cell-centered finite volume Computational Fluid Dynamics (CFD) software framework named PHengLEI which was originally developed for aerodynamics applications with arbitrary hybrid meshes. Overall, the newly developed GPU framework demonstrate up to 4.8 speedup comparing with 18 MPI tasks run on the latest Intel CPU node. Furthermore, the enormous efforts have been invested to optimize data dependencies which could lead to race condition due to unstructured mesh indirect addressing and related reduction math operations. With careful comparison between our optimised graph coloring and atomic operations using a series of numerical tests with different mesh sizes, the results show that atomic operations are more efficient than our optimised graph coloring in all of the test cases on Nvidia Tesla GPU V100. Specifically, for the summation operation, using atomicAdd is twice as fast as graph coloring. For the maximum operation, a speedup of 1.5 to 2 is found for atomicMax vs. graph coloring.

The birth of Bitcoin ushered in the era of cryptocurrency, which has now become a financial market attracted extensive attention worldwide. The phenomenon of startups launching Initial Coin Offerings (ICOs) to raise capital led to thousands of tokens being distributed on blockchains. Many studies have analyzed this phenomenon from an economic perspective. However, little is know about the characteristics of participants in the ecosystem. To fill this gap and considering over 80% of ICOs launched based on ERC20 token on Ethereum, in this paper, we conduct a systematic investigation on the whole Ethereum ERC20 token ecosystem to characterize the token creator, holder, and transfer activity. By downloading the whole blockchain and parsing the transaction records and event logs, we construct three graphs, namely token creator graph, token holder graph, and token transfer graph. We obtain many observations and findings by analyzing these graphs. Besides, we propose an algorithm to discover potential relationships between tokens and other accounts. The reported case shows that our algorithm can effectively reveal entities and the complex relationship between various accounts in the token ecosystem.

MOTIVATION: RNA secondary structure plays a vital role in fundamental cellular processes, and identification of RNA secondary structure is a key step to understand RNA functions. Recently, a few experimental methods were developed to profile genome-wide RNA secondary structure, i.e. the pairing probability of each nucleotide, through high-throughput sequencing techniques. However, these high-throughput methods have low precision and cannot cover all nucleotides due to limited sequencing coverage. RESULTS: Here, we have developed a new method for the prediction of genome-wide RNA secondary structure profile from RNA sequence based on the extreme gradient boosting technique. The method achieves predictions with areas under the receiver operating characteristic curve (AUC) >0.9 on three different datasets, and AUC of 0.888 by another independent test on the recently released Zika virus data. These AUCs are consistently >5% greater than those by the CROSS method recently developed based on a shallow neural network. Further analysis on the 1000 Genome Project data showed that our predicted unpaired probabilities are highly correlated (>0.8) with the minor allele frequencies at synonymous, non-synonymous mutations, and mutations in untranslated regions, which were higher than those generated by RNAplfold. Moreover, the prediction over all human mRNA indicated a consistent result with previous observation that there is a periodic distribution of unpaired probability on codons. The accurate predictions by our method indicate that such model trained on genome-wide experimental data might be an alternative for analytical methods. AVAILABILITY AND IMPLEMENTATION: The GRASP is available for academic use at https://github.com/sysu-yanglab/GRASP. SUPPLEMENTARY INFORMATION: Supplementary data are available online.

Summary As data analysis scenarios keep increasing on high‐performance computing systems, the ability to select a small fraction of data from a large volume of scientific data sets is vital to accelerate scientific discovery. However, parallel file systems lack the ability to provide efficient data locating services at the granularity of both a file and a record. Existing methods for identifying and indexing data are often domain‐specific and do not scale to large scientific data sets. In this paper, we describe the design and implementation of UniIndex framework, which combines our proposed techniques for user‐annotation extraction, in‐memory cache layer, in‐situ indexing, and parallel query processing. Acting as middleware on top of production file systems, UniIndex enables efficient data locating services with minimal user effort. Our evaluations show that UniIndex can locate target files from directories containing millions of files in microseconds. By applying in situ indexing and the lightweight range‐bitmap index, record‐level index building time can be dramatically reduced while maintaining up to two orders of magnitude query speedup than scanning the entire data set.

Summary Iterative linear algebra methods to solve linear systems and eigenvalue problems with non‐Hermitian matrices are important for both the simulation arising from diverse scientific fields and the applications related to big data, machine learning, and artificial intelligence. The spectral property of these matrices has impacts on the convergence of these solvers. Moreover, with the increase of the size of applications, iterative methods are implemented in parallel on clusters. Analysis of their behaviors with non‐Hermitian matrices on supercomputers is so complex that we need to generate large‐scale matrices with different given spectra for benchmarking. These test matrices should be non‐Hermitian and nontrivial, with high dimension. This paper highlights a scalable matrix generator that constructs large sparse matrices using the user‐defined spectrum, and the eigenvalues of generated matrices are ensured to be the same as the predefined spectrum. This generator is implemented on CPUs and multi‐GPUs platforms, with good strong and weak scaling performance on several supercomputers. We also propose a method to verify its ability to guarantee the given spectra. Finally, we give an example to evaluate the numerical properties and parallel performance of iterative methods using this matrix generator.

Molecular dynamics (MD) is a computer simulation method of studying physical movements of atoms and molecules that provide detailed microscopic sampling on molecular scale. With the continuous efforts and improvements, MD simulation gained popularity in materials science, biochemistry and biophysics with various application areas and expanding data scale. Assisted Model Building with Energy Refinement (AMBER) is one of the most widely used software packages for conducting MD simulations. However, the speed of AMBER MD simulations for system with millions of atoms in microsecond scale still need to be improved. In this paper, we propose a parallel acceleration strategy for AMBER on the Tianhe-2 supercomputer. The parallel optimization of AMBER is carried out on three different levels: fine grained OpenMP parallel on a single CPU, single node CPU/MIC parallel optimization and multi-node multi-MIC collaborated parallel acceleration. By the three levels of parallel acceleration strategy above, we achieved the highest speedup of 25-33 times compared with the original program.

Data analytics has become an integral part of large-scale scientific computing. Among various data analytics frameworks, MapReduce has gained the most traction. Although some efforts have been made to enable efficient MapReduce for supercomputing systems, they are often limited to fairly homogeneous workloads where equal partitioning of input data across tasks results in essentially equal output or temporary data generated on each task. For workloads that are more skewed, however, current implementations can result in imbalance in memory usage and, consequently, can cause a slowdown in execution time and a loss in data scalability. To tackle this problem, we enhance a previously published memory-conscious MapReduce over MPI framework called Mimir. Our enhancements to Mimir include combiner and dynamic repartition optimizations to minimize and balance memory usage and to achieve close to optimal balance of the memory usage across processes and to reduce the execution time by up to 12 times. Experimental results show that Mimir can scale to at least 3072 processes on the Tianhe-2 supercomputer on skewed datasets.

In order to make the query retrieve much more related products, some query rewrite methods have been proposed to obtain a set of candidate queries which can infer users' search intents and reduce the vocabulary gap between the original query and title of related products. However, previous studies ignore that some candidate queries may change users' search intents and retrieve irrelevant products. As a result, users' search experience will be impacted significantly. To reduce this influence, we need to design a semantic matching model to determine whether the candidate query change the original query's search intents (semantics). In addition, building a semantic matching model faces the following challenges: 1) Queries are usually very short and have limited information. It is very hard to learn an effective semantic matching model with the textual information of queries and candidate queries. 2) In order to get a generalized and effective mode, sufficient data samples are required to train the model. However, the cost of labeling is very huge. In order to address the above challenges, we propose an active and deep semantic matching framework (ActiveMatch) which is composed of two components. One component is the deep semantic matching (DSM) model which can make full use of the search log information to enhance the representation of queries and candidate queries. Then, it can estimate the semantic similarity between the original query and the candidate query more accurately. The other component is an uncertainty and novelty sampling (UNS) strategy which selects the samples to label based on the difficulty of the model estimating and the probability of the occurrence of new words. It not only reduces the cost of labeling but also ensures the effectiveness of the model. The experimental results on the Taobao e-commercial search platform verify the effectiveness of our framework.

The efficiency of metadata processing affects the file system performance significantly. There are two bottlenecks in metadata management in existing local file systems: 1) Path lookup is costly because it causes a lot of disk I/Os, which makes metadata operations inefficient. 2) Existing file systems have deep I/O stack in metadata management, resulting in additional processing overhead. To solve these two bottlenecks, we decoupled data and metadata management and proposed a metadata management layer for local file systems. First, we separated the metadata based on their locations in the namespace tree and aggregated the metadata into fixed-size metadata buckets (MDBs). This design fully utilizes the metadata locality and improves the efficiency of disk I/O in the path lookup. Second, we customized an efficient MDB storage system on the raw storage device. This design simplifies the file system I/O stack in the metadata management and allows metadata lookup to be completed with constant time complexity. Finally, this metadata management layer gives users the flexibility to choose metadata storage devices. We implemented a prototype called Otter. Our evaluation demonstrated that Otter outperforms native EXT4, XFS, Btrfs, BetrFS and TableFS in many metadata operations. For instance, Otter has 1.2 times to 9.6 times performance improvement over other tested file systems in file opening.

Video temporal action detection aims to temporally localize and recognize the action in untrimmed videos. Existing one-stage approaches mostly focus on unifying two subtasks, i.e., localization of action proposals and classification of each proposal through a fully shared backbone. However, such design of encapsulating all components of two subtasks in one single network might restrict the training by ignoring the specialized characteristic of each subtask. In this paper, we propose a novel Decoupled Single Shot temporal Action Detection (Decouple-SSAD) method to mitigate such problem by decoupling the localization and classification in a one-stage scheme. Particularly, two separate branches are designed in parallel to enable each component to own representations privately for accurate localization or classification. Each branch produces a set of action anchor layers by applying deconvolution to the feature maps of the main stream. Each branch produces a set of feature maps by applying deconvolution to the feature maps of the main stream. High-level semantic information from deeper layers is thus incorporated to enhance the feature representations. We conduct extensive experiments on THUMOS14 dataset and demonstrate superior performance over state-of-the-art methods. Our code is available online.

Analyzing large volumes of data is becoming more and more important in various scientific computing domains. MapReduce over MPI frameworks are an appealing solution to enable scalable big data analytics on supercomputing systems. These systems can further leverage features of MapReduce applications by merging (key/value) pairs before the reduce function in combiner optimizations. In this paper, we propose a pipeline combiner workflow and integrate it into Mimir, a cutting-edge implementation of Map Reduce over MPI. Our results with real datasets on the Tianhe-2 supercomputer prove that our pipeline combiner workflow can reduce memory usage up to 51% and improve the overall performance up to 61%.

BACKGROUND: Novel sequence motifs detection is becoming increasingly essential in computational biology. However, the high computational cost greatly constrains the efficiency of most motif discovery algorithms. RESULTS: In this paper, we accelerate MEME algorithm targeted on Intel Many Integrated Core (MIC) Architecture and present a parallel implementation of MEME called MIC-MEME base on hybrid CPU/MIC computing framework. Our method focuses on parallelizing the starting point searching method and improving iteration updating strategy of the algorithm. MIC-MEME has achieved significant speedups of 26.6 for ZOOPS model and 30.2 for OOPS model on average for the overall runtime when benchmarked on the experimental platform with two Xeon Phi 3120 coprocessors. CONCLUSIONS: Furthermore, MIC-MEME has been compared with state-of-arts methods and it shows good scalability with respect to dataset size and the number of MICs. Source code: https://github.com/hkwkevin28/MIC-MEME .

In the original version of this article, the abbreviation ‘OWDM’ was incorrectly defined. The phrase ‘orthogonal wavelength division multiplexing’ should all be changed to ‘one-way wave depth migration’.

Single Nucleotide Polymorphism (SNP) detection is a fundamental procedure of whole genome analysis. SOAPsnp, a classic tool for detection, would take more than one week to analyze one typical human genome, which limits the efficiency of downstream analyses. In this paper, we present mSNP, an optimized version of SOAPsnp, which leverages Intel Xeon Phi coprocessors for large-scale SNP detection. Firstly, we redesigned the essential data structures of SOAPsnp, which significantly reduces memory footprint and improves computing efficiency. Then we developed a coordinated parallel framework for a higher hardware utilization of both CPU and Xeon Phi. Also, we tailored the data structures and operations to utilize the wide VPU of Xeon Phi to improve data throughput. Last but not the least, we proposed a read-based window division strategy to improve throughput and obtain better load balance. mSNP is the first SNP detection tool empowered by Xeon Phi. We achieved a 38x single thread speedup on CPU, without any loss in precision. Moreover, mSNP successfully scaled to 4,096 nodes on Tianhe-2. Our experiments demonstrate that mSNP is efficient and scalable for large-scale human genome SNP detection.

Molecular docking is an important method in computational drug discovery. In large-scale virtual screening, millions of small drug-like molecules (chemical compounds) are compared against a designated target protein (receptor). Depending on the utilized docking algorithm for screening, this can take several weeks on conventional HPC systems. However, for certain applications including large-scale screening tasks for newly emerging infectious diseases such high runtimes can be highly prohibitive. In this paper, we investigate how the massively parallel neo-heterogeneous architecture of Tianhe-2 Supercomputer consisting of thousands of nodes comprising CPUs and MIC coprocessors that can efficiently be used for virtual screening tasks. Our proposed approach is based on a coordinated parallel framework called mD3DOCKxb in which CPUs collaborate with MICs to achieve high hardware utilization. mD3DOCKxb comprises a novel efficient communication engine for dynamic task scheduling and load balancing between nodes in order to reduce communication and I/O latency. This results in a highly scalable implementation with parallel efficiency of over 84% (strong scaling) when executing on 8,000 Tianhe-2 nodes comprising 192,000 CPU cores and 1,368,000 MIC cores.

K-mer counting is a fundamental operation in DNA research and genome analytics; its application includes estimating genome assembly, understanding similarities in genomic samples, and merging a newly processed genome with a reference genome. As the genome dataset becomes larger and larger, designing a highly optimized distributed-memory implementation becomes more and more important. Current distributed-memory solutions have two limitations: they have a high memory footprint, and they do not provide advanced optimizations for loading enormous genome datasets into memory. Based on these observations, we present Bloomfish, a distributed, memory-efficient, scalable solution to the limits of current work. To keep a low memory footprint, Bloomfish leverages the compact hash array design of the single-node Jellyfish system and the optimized workflow of the high-performance MapReduce framework Mimir. We have also codesigned Mimir's I/O to efficiently load enormous datasets. We ran Bloomfish on the Tianhe-2 supercomputer with large sequence datasets (up to 24 TB). Our results show that Bloomfish achieves unprecedented scalability in genome analytics.

In this paper we present Mimir, a new implementation of MapReduce over MPI. Mimir inherits the core principles of existing MapReduce frameworks, such as MR-MPI, while redesigning the execution model to incorporate a number of sophisticated optimization techniques that achieve similar or better performance with significant reduction in the amount of memory used. Consequently, Mimir allows significantly larger problems to be executed in memory, achieving large performance gains. We evaluate Mimir with three benchmarks on two highend platforms to demonstrate its superiority compared with that of other frameworks.

With supercomputers developing towards exascale, the number of compute cores increases dramatically, making more complex and larger-scale applications possible. The input/output (I/O) requirements of large-scale applications, workflow applications, and their checkpointing include substantial bandwidth and an extremely low latency, posing a serious challenge to high performance computing (HPC) storage systems. Current hard disk drive (HDD) based underlying storage systems are becoming more and more incompetent to meet the requirements of next-generation exascale supercomputers. To rise to the challenge, we propose a hierarchical hybrid storage system, on-line and near-line file system (ONFS). It leverages dynamic random access memory (DRAM) and solid state drive (SSD) in compute nodes, and HDD in storage servers to build a three-level storage system in a unified namespace. It supports portable operating system interface (POSIX) semantics, and provides high bandwidth, low latency, and huge storage capacity. In this paper, we present the technical details on distributed metadata management, the strategy of memory borrow and return, data consistency, parallel access control, and mechanisms guiding downward and upward migration in ONFS. We implement an ONFS prototype on the TH-1A supercomputer, and conduct experiments to test its I/O performance and scalability. The results show that the bandwidths of single-thread and multi-thread ‘read’/‘write’ are 6-fold and 5-fold better than HDD-based Lustre, respectively. The I/O bandwidth of data-intensive applications in ONFS can be 6.35 times that in Lustre.

Molecular dynamics (MD) is a computer simulation method of studying physical movements of atoms and molecules that provide detailed microscopic sampling on molecular scale. With the continuous efforts and improvements, MD simulation gained popularity in materials science, biochemistry and biophysics with various application areas and expanding data scale. Assisted Model Building with Energy Refinement (AMBER) is one of the most widely used software packages for conducting MD simulations. However, the speed of AMBER MD simulations for system with millions of atoms in microsecond scale still need to be improved. In this paper, we propose a parallel acceleration strategy for AMBER on Tianhe-2 supercomputer. The parallel optimization of AMBER is carried out on three different levels: fine grained OpenMP parallel on a single MIC, single-node CPU/MIC collaborated parallel optimization and multi-node multi-MIC collaborated parallel acceleration. By the three levels of parallel acceleration strategy above, we achieved the highest speedup of 25-33 times compared with the original program. Source Code: https://github.com/tianhe2/mAMBER.

Hard Disk Drive (HDD) based storage systems in high performance computing (HPC), such as Lustre, are being confronted with great challenges from massive parallel data-intensive applications, and becomes one of the most critical factors that influence applications performance. In this paper, we analyze the performance of Lustre, and statistically examine the distribution of file sizes, I/O patterns for typical data-intensive applications, as well as the utilization of CPU and memory resources in TH-1A supercomputer from the last five years. We present masFS, a novel file system for HPC that exploits available memory and SSD resources on compute nodes with little interference to applications running on the nodes. The masFS supports POSIX interface with an unique namespace and provides comprehensive file service for applications. We design and implement masFS with memory and SSD, and have deployed and evaluated it on TH-1A. Experimental results show that masFS works as general storage system and can meet the requirement of variety of I/O loads. It can achieve a speedup of 14.5x for read and 8.8x for write respectively in running benchmarks, and an average 7.6x faster I/O time in a real-world data-intensive application compared to Lustre.

Numerical simulation of thermal convection in the Earth's outer core requires extreme-scale computing due to the large temporal and spatial disparity, extreme physical parameters, rapid rotation and spherical geometry. In this work, the numerical simulation of the thermal convection in the Earth's outer core for CPU-MIC heterogeneous many-core systems is studied. Firstly, starting from a legacy parallel code based on the PETSc software package, a framework of the numerical simulation built on CPU-MIC heterogeneous many-core systems has been developed. Secondly, a sparse linear solver for CPUMIC heterogeneous many-core systems, which focuses on solving the two linear systems of the simulation, is presented and optimized. Thirdly, some computational kernels of the simulation, including sparse matrix-vector multiplication (SpMV) and polynomial preconditioner on distributed memory Xeon Phiaccelerated systems are implemented and optimized. In addition, in order to reduce the cost of data movement, we use methods to minimize the memory access, the PCI-E data transfer, and the MPI communication. Finally, some optimized measures are taken to the extended code. Experiments on Tianhe-2 Supercomputer show that as compared to the original code, our Xeon Phiaccelerated design is able to deliver 6.93x and 6.00x speedups for single MIC device and 64 MIC devices, respectively.

NoSQL systems are deployed as the core components for delivering big data Web services today. With growing main memory capacity, we witness the growing interest and deployment of in-memory NoSQL services (IM-NoSQL), which are designed to maximize the utilization of DRAM for ultra low latency services. To address the volatility of DRAM for in-memory computing services, persistence and failure recovery are important functionality for IM-NoSQL. In this paper we report an extensive measurement study on the performance of persistence and recovery for IM-NoSQL. We evaluate the performance and effectiveness of several common mechanisms used for persistence and recovery in the presence of server crashes, such as snapshot and logging based approaches. Through this study, we are able to answer some of the most frequently asked questions in provisioning of IM-NoSQL services: (i) Can an IM-NoSQL system work effectively when the available memory is insufficient to load the whole dataset? (ii) What is the overhead of maintaining snapshot compared to logging? (iii) How fast an IM-NoSQL system can recover in the presence of failure? And (iv) how does an IM-NoSQL system respond to the different persistence models? We report our comprehensive measurement results on execution, persistence and recovery performance of Redis, a representative implementation of IM-NoSQL services.

Solid State Drives (SSDs) have been extensively deployed as the cache of hard disk-based storage systems. The SSD-based cache generally supplies ultra-large capacity, whereas managing so large a cache introduces excessive memory overhead, which in turn makes the SSD-based cache neither cost-effective nor energy-efficient. This work targets to reduce the memory overhead introduced by the replacement policy of SSD-based cache. Traditionally, data structures involved in cache replacement policy reside in main memory. While these in-memory data structures are not suitable for SSD-based cache any more since the cache is much larger than ever. We propose a memory-efficient framework which keeps most data structures in SSD while just leaving the memory-efficient data structure (i.e., a new bloom proposed in this work) in main memory. Our framework can be used to implement any LRU-based replacement policies under negligible memory overhead. We evaluate our proposals via theoretical analysis and prototype implementation. Experimental results demonstrate that, our framework is practical to implement most replacement policies for large caches, and is able to reduce the memory overhead by about$10 \times$.

Heterogeneous platforms composed of multi-core CPUs and different types of accelerators, like GPUs and Xeon Phi, are becoming popular for data parallel applications. The heterogeneity of the hardware mix and the diversity of the applications pose significant challenges to exploiting such platforms. In this situation, an effective workload partitioning between processing units is critically important for improving application performance. This partitioning is a function of the hardware capabilities as well as the application and the dataset to be used. In this work, we present a systematic approach to solve the partitioning problem. Specifically, we use modeling, profiling, and prediction techniques to quickly and correctly predict the optimal workload partitioning and the right hardware configuration to use. Our approach effectively characterizes the platform heterogeneity, efficiently determines the accurate partitioning, and easily adapts to new platforms, different application types, and different datasets. Experimental evaluation on 13 applications shows that our approach delivers excellent performance improvement of 1.2$\times$–14.6$\times$over a single-processor execution, and accurate partitioning with in most cases below 10 percent performance gap versus an oracle-based partitioning.

Molecular docking is a time consuming process, and it requires a substantial amount of computing power. D3DOCkxb was developed for investigating the effects of halogen bond in drug discovery by adding two precise score functions to Auto Dock. The docking accuracy of D3DOCkxb is better than Auto Dock, which can be attributed to a more complicated processing logic of D3DOCkxb. Consequently, it is an even more challenging task to do parallel optimization on D3DOCkxb. In this paper, we developed mD3DOCkxb, a MIC enabled version of D3DOCkxb, which utilizes Intel Xeon Phi, a Many-Integrated Core (MIC) accelerator, to boost the docking performance. We parallelized the Lamarckian Genetic Algorithm (LGA) in D3DOCKxb with OpenMP and port it to MIC with a number of optimization. And 12x to 18x speedup can be achieved, depending on the number of LGA iterations.

Graph structure which is often used to model the relationship between the data items has drawn more and more attention. The graph datasets from many important domains have the property called scale-free. In the scale-free graphs, there exist the hubs, which have much larger degree than the average value. The hubs may cause the problems of load imbalance, poor scalability and high communication overhead when the graphs are processed in the distributed memory systems. In this paper, we design an asynchronous graph processing framework targeted for distributed memory by considering the hubs as a separate part of the vertexes, which we call it the hub-centric idea. Specifically speaking, a hub-duplicate graph partitioning method is proposed to balance the workload and reduce the communication overhead. At the same time, an efficient asynchronous state synchronization method for the duplicates is also proposed. In addition, a priority scheduling strategy is applied to further reduce the communication overhead.

Whole genome re-sequencing plays a crucial role in biomedical studies. The emergence of genomic big data calls for an enormous amount of computing power. However, current computational methods are inefficient in utilizing available computational resources. In this paper, we address this challenge by optimizing the utilization of the fastest supercomputer in the world - TH-2 supercomputer. TH-2 is featured by its neo-heterogeneous architecture, in which each compute node is equipped with 2 Intel Xeon CPUs and 3 Intel Xeon Phi coprocessors. The heterogeneity and the massive amount of data to be processed pose great challenges for the deployment of the genome analysis software pipeline on TH-2. Runtime profiling shows that SOAP3-dp and SOAPsnp are the most time-consuming components (up to 70% of total runtime) in a typical genome-analyzing pipeline. To optimize the whole pipeline, we first devise a number of parallel and optimization strategies for SOAP3-dp and SOAPsnp, respectively targeting each node to fully utilize all sorts of hardware resources provided both by CPU and MIC. We also employ a few scaling methods to reduce communication between different nodes. We then scaled up our method on TH-2. With 8192 nodes, the whole analyzing procedure took 8.37 hours to finish the analysis of a 300 TB dataset of whole genome sequences from 2,000 human beings, which can take as long as 8 months on a commodity server. The speedup is about 700x.

Wayside supercapacitor based energy storage system (ESS) for urban rail transit can reuse the braking energy and reduce energy consumption. Most ESSs switch over from charging to discharging states based on voltage of DC bus. But open circuit voltage of DC power supply in subway system always changes, which would influence the normal operation of ESS. This paper will show the output characteristic of substation and the basic structure of supercapacitor based ESS. At the same time, influences of substation output voltage fluctuation on ESS will be analyzed. And then, a control strategy for eliminating the influence will be come up with. Finally, this paper will present the experiment results of 200kW supercapacitor based ESS prototype field test in Beijing subway.

Each cell of flash memory only survives a nominally given number of write/erasure cycles. Beyond the nominal lifespan, flash memory can still record digital information but the bit error rate increases rapidly with the increment of write/erase cycles. Erasure code is a conventional method used to recover corrupted data, but its redundant data produce a large number of additional writes, making the erasure code seem to be unsuitable for the write-sensitive flash memory. We argue that, erasure code influences the lifespan of flash memory in two conflicting directions: its inherent error correction capability enables the flash memory to survive beyond the nominal lifespan, while its redundant data wear out the lifespan of flash memory by increasing the write/erase cycles. This paper builds a theoretical model to analyze both the two aspects and demonstrates that the erasure code is able to extend the nominal lifespan of flash memory by as many as 30×.

In this work an ultra-scalable algorithm is designed and optimized to accelerate a 3D compressible Euler atmospheric model on the CPU-MIC hybrid system of Tianhe-2. We first reformulate the mesocale model to avoid long-latency operations, and then employ carefully designed inter-node and intra-node domain decomposition algorithms to achieve balance utilization of different computing units. Proper communication-computation overlap and concurrent data transfer methods are utilized to reduce the cost of data movement at scale. A variety of optimization techniques on both the CPU side and the accelerator side are exploited to enhance the in-socket performance. The proposed hybrid algorithm successfully scales to 6,144 Tianhe-2 nodes with a nearly ideal weak scaling efficiency, and achieve over 8 percent of the peak performance in double precision. This ultra-scalable hybrid algorithm may be of interest to the community to accelerating atmospheric models on increasingly dominated heterogeneous supercomputers.

In this paper, we propose a hybrid algorithm to enable and accelerate the High Performance Conjugate Gradient (HPCG) benchmark on a heterogeneous node with an arbitrary number of accelerators. In the hybrid algorithm, each subdomain is assigned to a node after a three-dimensional domain decomposition. The subdomain is further divided to several regular inner blocks and an outer part with a flexible inner-outer partitioning strategy. Each inner task is assigned to a MIC device and the size is adjustable to adapt the accelerator's computational power. The only outer part is assigned to CPU and the thickness of boundary size is also adjustable to maintain load balance between CPU and MICs. By properly fusing the computational kernels with preceding ones, we present an asynchronous data transfer scheme to better overlap local computation with the PCI-express data transfer. All basic HPCG kernels, especially the time-consuming sparse matrix-vector multiplication (SpMV) and the symmetric Gauss-Seidel relaxation (SymGS), are extensively optimized for both CPU and MIC, on both algorithmic and architectural levels. On a single node of Tianhe-2 which is composed of an Intel Xeon processor and three Intel Xeon Phi coprocessors, we successfully obtain an aggregated performance of 50.2 Gflops, which is around 1.5% of the peak performance.

Developing an efficient and highly scalable ray tracer for the Metropolis light transport algorithm is becoming increasingly important as the request for photorealistic images becomes a common trend. Although the Metropolis light transport algorithm has produced some of the most realistic images to date, it usually takes a great amount of time to render an image. The development of an efficient and highly scalable ray tracer for the Metropolis light transport algorithm is hard due in large part to the irregular memory access patterns, the imbalanced workload of light-carrying paths and the complicated mathematical model and complex physical processes. In this paper, we present a highly scalable physically based parallel ray tracer for the Metropolis light transport algorithm. Firstly, we present the idea of snapshot and sub-snapshot, then propose a novel assignment partitioning algorithm for compute nodes and CPU cores since the demand-driven assignment partitioning algorithms don't work. Secondly, we propose a physically based parallel ray racing framework for the Metropolis light transport algorithm, which is based on a master-worker architecture. Finally, we discuss the issue of granularity of the assignment partitioning and some optimization strategies for improving overall performance, then a hybrid scheduling strategy combining a static and dynamic scheduling strategy is described. Experiments show that our physically based ray tracer almost reaches linear speedup by using 26,400 CPU cores on the Tianhe-2 supercomputer. Our ray tracer is more efficient and highly scalable.

Although GPUs are considered ideal to accelerate massively data-parallel applications, there are still exceptions to this rule. For example, imbalanced applications cannot be efficiently processed by GPUs: despite the massive data parallelism, a varied computational workload per data point remains GPU-unfriendly. To efficiently process imbalanced applications, we exploit the use of heterogeneous platforms (GPUs and CPUs) by partitioning the workload to fit the usage patterns of the processors. In this work, we present our flexible and adaptive method that predicts the optimal partitioning. Our method aims to match a quantitative model of the application with the hardware capabilities of the platform, and calculates the optimal match according to a user-given criterion. We evaluate our method in terms of overall performance gain, prediction accuracy, flexibility and adaptivity. Our results, gathered from both synthetic and real-world workloads, show performance gains of up to 60%, accurate predictions for more than 90% of all the 1395 imbalanced workloads we have tested, and confirm that the method adapts correctly to application, dataset, and platform changes (both hardware and software). We conclude that model-based prediction of workload partitioning for heterogeneous platforms is feasible and useful for performance improvement.

Summary form only given: Since scalability is one of the major challenges for advanced HPC systems in the post-petascale and exascale era, innovative integrated technology designs are needed for new architecture as well as associated software stacks. We need to explore the capability of cpu, accelerator, interconnection, I/O storage system, and till whole system. This talk will discuss the way of scalability-centric HPC system hardware and software design related to the computation, communication, data procession, and fault tolerance. The experiences on the design and implementation of Tianhe systems will also be given. Furthermore, some investigations on architecture and software design for the next generation HPC system will be presented. In general, a co-design approach should be followed throughout the research and development activities to deliver a whole system for scalable computing, to support the large-scale domain applications efficiently.

This paper presents a hybrid algorithm for the petascale global simulation of atmospheric dynamics on Tianhe-2, the world's current top-ranked supercomputer developed by China's National University of Defense Technology (NUDT). Tianhe-2 is equipped with both Intel Xeon CPUs and Intel Xeon Phi accelerators. A key idea of the hybrid algorithm is to enable flexible domain partition between an arbitrary number of processors and accelerators, so as to achieve a balanced and efficient utilization of the entire system. We also present an asynchronous and concurrent data transfer scheme to reduce the communication overhead between CPU and accelerators. The acceleration of our global atmospheric model is conducted to improve the use of the Intel MIC architecture. For the single-node test on Tianhe-2 against two Intel Ivy Bridge CPUs (24 cores), we can achieve 2.07×, 3.18×, and 4.35× speedups when using one, two, and three Intel Xeon Phi accelerators respectively. The average performance gain from SIMD vectorization on the Intel Xeon Phi processors is around 5× (out of the 8× theoretical case). Based on successful computation-communication overlapping, large-scale tests indicate that a nearly ideal weak-scaling efficiency of 93.5% is obtained when we gradually increase the number of nodes from 6 to 8,664 (nearly 1.7 million cores). In the strong-scaling test, the parallel efficiency is about 77% when the number of nodes increases from 1,536 to 8,664 for a fixed 65,664 × 5,664 × 6 mesh with 77.6 billion unknowns.

A new sparse high performance conjugate gradient benchmark (HPCG) has been recently released to address challenges in the design of sparse linear solvers for the next generation extreme-scale computing systems. Key computation, data access, and communication pattern in HPCG represent building blocks commonly found in today's HPC applications. While it is a well known challenge to efficiently parallelize Gauss-Seidel smoother, the most time-consuming kernel in HPCG, our algorithmic and architecture-aware optimizations deliver 95% and 68% of the achievable bandwidth on Xeon and Xeon Phi, respectively. Based on available parallelism, our Xeon Phi shared-memory implementation of Gauss-Seidel smoother selectively applies block multi-color reordering. Combined with MPI parallelization, our implementation balances parallelism, data access locality, CG convergence rate, and communication overhead. Our implementation achieved 580 TFLOPS (82% parallelization efficiency) on Tianhe-2 system, ranking first on the most recent HPCG list in July 2014. In addition, we demonstrate that our optimizations not only benefit HPCG original dataset, which is based on structured 3D grid, but also a wide range of unstructured matrices.

Fault resilience has became a major issue for HPC systems, in particular in the perspective of future E-scale systems, which will consist of millions of CPU cores and other components. Fault tolerant MPI was proposed to offer support of software level fault tolerance approaches. However, the widely used MPI implementations, such as MPICH and Mvapich2, provide limited support for fault tolerance. This paper proposes NR-MPI, a Non-stop and Fault Resilient MPI. NR-MPI implements the semantics of FT-MPI based on MPICH. Specifically, this paper focuses on failure detection in MPI library, online failure recovery of communicators for multiple failures, friendly programming interface extending for NR-MPI. Furthermore, to support failure recovery of applications, NR-MPI implements data backup and restore interfaces based on double in-memory checkpoint/restart. We conduct experiments with NPB benchmarks on TH-1A supercomputer. Experimental results show that NR-MPI based fault tolerant programs can recover from failures online without restarting, and the overhead is small even for applications with tens of thousands of cores.

Developing highly scalable algorithms for global atmospheric modeling is becoming increasingly important as scientists inquire to understand behaviors of the global atmosphere at extreme scales. Nowadays, heterogeneous architecture based on both processors and accelerators is becoming an important solution for large-scale computing. However, large-scale simulation of the global atmosphere brings a severe challenge to the development of highly scalable algorithms that fit well into state-of-the-art heterogeneous systems. Although successes have been made on GPU-accelerated computing in some top-level applications, studies on fully exploiting heterogeneous architectures in global atmospheric modeling are still very less to be seen, due in large part to both the computational difficulties of the mathematical models and the requirement of high accuracy for long term simulations.

This paper introduces an extensible distributed file system framework, YaFS, using heterogeneous online storage services as its back-ends. It provides a configurable solution for simplifying the usage of multiple storage resources and accessing data ubiquitously and safely. YaFS is POSIX compliant, so that it could support most of the existing applications seamlessly. An offline mode is used to cope with the challenged unreliable network environment. We implement a storage abstraction layer and a plug-in mechanism for uniformly accessing different storage services transparently, and it makes the system expanded easily. YaFS could effectively support for storing large object to size limited services and achieving high aggregate bandwidth by striping data on multiple servers with bandwidth-saving method. The evaluation on a prototype implementation with email services as its storage back-end shows that the performance and usability of the framework is viable.