Long-context large language models (LLMs) have seen widespread adoption in recent years. However, during inference, the key-value (KV) cache—which stores intermediate activations—consumes significant memory, particularly as sequence lengths grow. Quantization offers a promising path to compress KV cache, but existing 2-bit approaches fall short of achieving optimal inference efficiency due to hardware-unfriendly algorithms and system implementations.

We explore Generalizable Tumor Segmentation, aiming to train a single model for zero-shot tumor segmentation across diverse anatomical regions. Existing methods face limitations related to segmentation quality, scalability, and the range of applicable imaging modalities. In this paper, we uncover the potential of the internal representations within frozen medical foundation diffusion models as highly efficient zero-shot learners for tumor segmentation by introducing a novel framework named DiffuGTS. DiffuGTS creates anomaly-aware open-vocabulary attention maps based on text prompts to enable generalizable anomaly segmentation without being restricted by a predefined training category list. To further improve and refine anomaly segmentation masks, DiffuGTS leverages the diffusion model, transforming pathological regions into high-quality pseudo-healthy counterparts through latent space inpainting, and applies a novel pixel-level and feature-level residual learning approach, resulting in segmentation masks with significantly enhanced quality and generalization. Comprehensive experiments on four datasets and seven tumor categories demonstrate the superior performance of our method, surpassing current state-of-the-art models across multiple zero-shot settings. Codes are available at https://github.com/Yankai96/DiffuGTS.

Developing agents capable of navigating to a target location based on language instructions and visual information, known as vision-language navigation (VLN), has attracted widespread interest. Most research has focused on ground-based agents, while UAV-based VLN remains relatively underexplored. Recent efforts in UAV vision-language navigation predominantly adopt ground-based VLN settings, relying on predefined discrete action spaces and neglecting the inherent disparities in agent movement dynamics and the complexity of navigation tasks between ground and aerial environments. To address these disparities and challenges, we propose solutions from three perspectives: platform, benchmark, and methodology. To enable realistic UAV trajectory simulation in VLN tasks, we propose the OpenUAV platform, which features diverse environments, realistic flight control, and extensive algorithmic support. We further construct a target-oriented VLN dataset consisting of approximately 12k trajectories on this platform, serving as the first dataset specifically designed for realistic UAV VLN tasks. To tackle the challenges posed by complex aerial environments, we propose an assistant-guided UAV object search benchmark called UAV-Need-Help, which provides varying levels of guidance information to help UAVs better accomplish realistic VLN tasks. We also propose a UAV navigation LLM that, given multi-view images, task descriptions, and assistant instructions, leverages the multimodal understanding capabilities of the MLLM to jointly process visual and textual information, and performs hierarchical trajectory generation. The evaluation results of our method significantly outperform the baseline models, while there remains a considerable gap between our results and those achieved by human operators, underscoring the challenge presented by the UAV-Need-Help task.

The group of permutations $S_n$, also known as the finite symmetric groups, are essential in fields such as combinatorics, physics, and chemistry. However, learning a probability distribution over $S_n$ poses significant challenges due to its intractable size and discrete nature. In this paper, we introduce *SymmetricDiffusers*, a novel discrete diffusion model that simplifies the task of learning a complicated distribution over $S_n$ by decomposing it into learning simpler transitions of the reverse diffusion using deep neural networks. We identify the riffle shuffle as an effective forward transition and provide empirical guidelines for selecting the diffusion length based on the theory of random walks on finite groups. Additionally, we propose a generalized Plackett-Luce (PL) distribution for the reverse transition, which is provably more expressive than the PL distribution. We further introduce a theoretically grounded "denoising schedule" to improve sampling and learning efficiency. Extensive experiments show that our model achieves state-of-the-art or comparable performance on solving tasks including sorting 4-digit MNIST images, jigsaw puzzles, and traveling salesman problems. Our code is released at <https://github.com/DSL-Lab/SymmetricDiffusers>.

Unmanned Aerial Vehicles (UAVs) are evolving into language-interactive platforms, enabling more intuitive forms of human-drone interaction. While prior works have primarily focused on high-level planning and long-horizon navigation, we shift attention to language-guided fine-grained trajectory control, where UAVs execute short-range, reactive flight behaviors in response to language instructions. We formalize this problem as the Flying-on-a-Word (Flow) task and introduce UAV imitation learning as an effective approach. In this framework, UAVs learn fine-grained control policies by mimicking expert pilot trajectories paired with atomic language instructions. To support this paradigm, we present UAV-Flow, the first real-world benchmark for language-conditioned, fine-grained UAV control. It includes a task formulation, a large-scale dataset collected in diverse environments, a deployable control framework, and a simulation suite for systematic evaluation. Our design enables UAVs to closely imitate the precise, expert-level flight trajectories of human pilots and supports direct deployment without sim-to-real gap. We conduct extensive experiments on UAV-Flow, benchmarking VLN and VLA paradigms. Results show that VLA models are superior to VLN baselines and highlight the critical role of spatial grounding in the fine-grained Flow setting.

The Mixture of Experts (MoE) architecture enhances model quality by scaling up model parameters. However, its development is hindered in distributed training scenarios due to significant communication overhead and expert load imbalance. Existing methods, which only allow for coarse-grained overlapping of communication and computation, slightly alleviate communication costs but at the same time, they introduce a notable impairment of computational efficiency. Furthermore, current approaches to addressing load imbalance often compromise model quality.

Micro-scaling General Matrix Multiplication (MX-GEMM), which leverages 8-bit micro-scaling format (MX-format) inputs, represents a significant step forward in accelerating deep learning workloads. The MX-format space is diverse, encompassing various scaling patterns and granularities. However, current MX-GEMM implementations typically adopt a model-oriented approach, where format customization is tailored to individual models. This results in three key limitations: rigid problem-kernel coupling, inefficient promotion operations, and overlooked quantization overhead.

General matrix-matrix multiplication (GEMM) is a fundamental operation in both deep learning and scientific computing. To accelerate these workloads, GPUs with a large number of streaming multiprocessors (SMs) are widely used. However, as modern GPUs scale in core count and adopt larger tile sizes, the wave quantization problem induced by partially filled waves results in growing hardware underutilization and substantially degraded performance. Existing solutions to this problem often suffer from low execution efficiency or introduce additional synchronization overhead.

Domain shifts such as sensor type changes and geographical situation variations are prevalent in Autonomous Driving (AD), which poses a challenge since AD model relying on the previous domain knowledge can be hardly directly deployed to a new domain without additional costs. In this paper, we provide a new perspective and approach of alleviating the domain shifts, by proposing a Reconstruction-Simulation-Perception (ReSimAD) scheme. Specifically, the implicit reconstruction process is based on the knowledge from the previous old domain, aiming to convert the domain-related knowledge into domain-invariant representations, e.g., 3D scene-level meshes. Besides, the point clouds simulation process of multiple new domains is conditioned on the above reconstructed 3D meshes, where the target-domain-like simulation samples can be obtained, thus reducing the cost of collecting and annotating new-domain data for the subsequent perception process. For experiments, we consider different cross-domain situations such as Waymo-to-KITTI, Waymo-to-nuScenes, etc, to verify the zero-shot target-domain perception using ReSimAD. Results demonstrate that our method is beneficial to boost the domain generalization ability, even promising for 3D pre-training. Code and simulated points are available at: https://github.com/PJLab-ADG/3DTrans

Data augmentation methods to enhance perception performance in adverse weather have recently attracted considerable attention. Most of the LiDAR data augmentation methods post-process the existing dataset by physics-based models or machine-learning methods. However, due to the limited environmental annotations and the fixed vehicle trajectories in existing datasets, it is challenging to edit the scene and expand the diversity of traffic flow and scenario. To this end, we propose a simulator-based physical modeling approach to augment LiDAR data in rainy weather, enhancing the performance of the perception model. We complete the modeling task of the rainy weather effect in the CARLA simulator and establish a data collection pipeline for LiDAR. Furthermore, we pay special attention to the spray generated by vehicles in rainy weather and simulate this phenomenon through the Spray Emitter method we developed. In addition, considering the influence of different weather conditions on point cloud intensity, we develop a prediction network to forecast the intensity of the LiDAR echo. This enables us to complete the rainy weather simulation of 4D point cloud data. In the experiment, we observe that the model augmented by our synthetic dataset improves the performance for 3D object detection in rainy weather. Both code and dataset are available at https://github.com/PJLab-ADG/PCSim#rainypcsim.

Deep Learning Recommendation Models (DLRMs) are pivotal in various sectors, yet they are hindered by the high memory demands of embedding tables and the significant communication overhead in distributed training environments. Traditional approaches, like Tensor-Train (TT) decomposition, although effective for compressing these tables, introduce substantial computational burdens. Furthermore, existing frameworks for distributed training are inadequate due to the excessive data exchange requirements.This paper proposes EcoRec, an advanced library designed to expedite the training of DLRMs through a synergistic integration of TT decomposition technology and distributed training. EcoRec introduces a novel computation pattern that eliminates redundancy in TT operations, alongside an efficient multiplication pathway, significantly reducing computational time. Additionally, it provides a unique micro-batching technique with sorted indices to decrease memory demands without additional computational costs. EcoRec also features a novel pipeline training system for embedding layers, ensuring balanced data distribution and enhanced communication efficiency. EcoRec, built on PyTorch and CUDA, has been evaluated on a 32 GPU cluster. The results show EcoRec significantly outperforms the existing ELRec system, achieving up to a $3.1 \times$ speedup and a 38.5% reduction in memory requirements. EcoRec marks a notable advancement in high-performance DLRM training.

Efficient training of Graph Neural Networks (GNNs) on billion-edge graphs poses significant challenges due to memory constraints and data transfer bottlenecks, particularly affecting GPU-based sampling. Traditional methods either face severe CPU-GPU data transfer bottlenecks or encounter excessive data shuffling and synchronization overheads in multi-GPU setups. To overcome these challenges in GNN training on large-scale graphs, we introduce HyDRA, a pioneering framework that elevates mini-batch, sampling-based training. HyDRA innovates in multi-GPU memory sharing and multi-node feature retrieval, transforming cross-GPU sampling by seamlessly integrating sampling and data transfer into a single kernel operation. It develops a hybrid pointer-driven data placement technique to enhance neighbor retrieval efficiency, designs a targeted replication strategy for high-degree vertices to reduce communication overhead, and leverages dynamic cross-batch data orchestration with pipelining to minimize redundant data transfers. Evaluated on systems equipped with up to 64 A100 GPUs, HyDRA significantly outperforms current leading methods, achieving $1.4 x$ to 5.3x faster training speeds compared to DSP and DGL-UVA and demonstrating up to a 42x improvement in multi-GPU scalability. HyDRA sets a new benchmark for high-performance GNN training at large scales.

Operator fusion, a key technique to improve data locality and alleviate GPU memory bandwidth pressure, often fails to extend to the fusion of multiple compute-intensive operators due to saturated computation throughput. However, the dynamicity of tensor dimension sizes could potentially lead to these operators becoming memory-bound, necessitating the generation of fused kernels — a task hindered by limited search spaces for fusion strategies, redundant memory access, and prolonged tuning time, leading to sub-optimal performance and inefficient deployment. We introduce MCFuser, a pioneering framework designed to overcome these obstacles by generating high-performance fused kernels for what we define as memory-bound compute-intensive (MBCI) operator chains. Leveraging high-level tiling expressions to delineate a comprehensive search space, coupled with Directed Acyclic Graph (DAG) analysis to eliminate redundant memory accesses, MCFuser streamlines kernel optimization. By implementing guidelines to prune the search space and incorporating an analytical performance model with a heuristic search, MCFuser not only significantly accelerates the tuning process but also demonstrates superior performance. Benchmarked against leading compilers like Ansor on NVIDIA A100 and RTX3080 GPUs, MCFuser achieves up to a 5.9x speedup in kernel performance and outpaces other baselines while reducing tuning time by over $\mathbf{7 0}$-fold, showcasing its agility.

Transformer-based models have made significant advancements across various domains, largely due to the self-attention mechanism’s ability to capture contextual relationships in input sequences. However, processing long sequences remains computationally expensive for Transformer models, primarily due to theO(n2) complexity associated with self-attention. To address this, sparse attention has been proposed to reduce the quadratic dependency to linear. Nevertheless, deploying the sparse transformer efficiently encounters two major obstacles: 1) Existing system optimizations are less effective for the sparse transformer due to the algorithm’s approximation properties leading to fragmented attention, and 2) the variability of input sequences results in computation and memory access inefficiencies. We present Raptor-T, a cutting-edge transformer framework designed for handling long and variable-length sequences. Raptor-T harnesses the power of the sparse transformer to reduce resource requirements for processing long sequences while also implementing system-level optimizations to accelerate inference performance. To address the fragmented attention issue, Raptor-T employs fused and memory-efficient Multi-Head Attention. Additionally, we introduce an asynchronous data processing method to mitigate GPU-blocking operations caused by sparse attention. Furthermore, Raptor-T minimizes padding for variable-length inputs, effectively reducing the overhead associated with padding and achieving balanced computation on GPUs. In evaluation, we compare Raptor-T’s performance against state-of-the-art frameworks on an NVIDIA A100 GPU. The experimental results demonstrate that Raptor-T outperforms FlashAttention-2 and FasterTransformer, achieving an impressive average end-to-end performance improvement of 3.41X and 3.71X, respectively.

Recently, Temporal Graph Neural Networks (TGNNs), as an extension of Graph Neural Networks, have demonstrated remarkable effectiveness in handling dynamic graph data. Distributed TGNN training requires efficiently tackling temporal dependency, which often leads to excessive cross-device communication that generates significant redundant data. However, existing systems are unable to remove the redundancy in data reuse and transfer, and suffer from severe communication overhead in a distributed setting. This work introduces Sven, a co-designed algorithm-system library aimed at accelerating TGNN training on a multi-GPU platform. Exploiting dependency patterns of TGNN models, we develop a redundancy-free graph organization to mitigate redundant data transfer. Additionally, we investigate communication imbalance issues among devices and formulate the graph partitioning problem as minimizing the maximum communication balance cost, which is proved to be an NP-hard problem. We propose an approximation algorithm called Re-FlexBiCut to tackle this problem. Furthermore, we incorporate prefetching, adaptive micro-batch pipelining, and asynchronous pipelining to present a hierarchical pipelining mechanism that mitigates the communication overhead. Sven represents the first comprehensive optimization solution for scaling memory-based TGNN training. Through extensive experiments conducted on a 64-GPU cluster, Sven demonstrates impressive speedup, ranging from 1.9x to 3.5x, compared to State-of-the-Art approaches. Additionally, Sven achieves up to 5.26x higher communication efficiency and reduces communication imbalance by up to 59.2%.

In recent years, the Mixture-of-Experts (MoE) technique has gained widespread popularity as a means to scale pretrained models to exceptionally large sizes. Dynamic activation of experts allows for conditional computation, increasing the number of parameters of neural networks, which is critical for absorbing the vast amounts of knowledge available in many deep learning areas. However, despite the existing system and algorithm optimizations, there are significant challenges to be tackled when it comes to the inefficiencies of communication and memory consumption. In this paper, we present the design and implementation of MPMoE, a high-performance library that accelerates MoE training with adaptive and memory-efficient pipeline parallelism. Inspired by that the MoE training procedure can be divided into multiple independent sub-stages. We design a pipeline parallelism method for reducing communication latency by overlapping with computation operations. Further, we analyze the memory footprint breakdown of MoE training and identify that activations and temporary buffers are the primary contributors to the overall memory footprint. Toward memory efficiency, we propose memory reuse strategies to reduce memory requirements by eliminating memory redundancies. Finally, to optimize pipeline granularity and memory reuse strategies jointly, we propose a profile-based algorithm and a performance model to determine the configurations of MPMoE at runtime. We implement MPMoE upon PyTorch and evaluate it with common MoE models in two physical clusters, including 64 NVIDIA A100 GPU cards and 16 NVIDIA V100 GPU cards. Compared with the state-of-art approach, MPMoE achieves up to 2.3× speedup while reducing more than 30% memory footprint for training large models.

Temporal Graph Neural Networks(TGNNs) extend the success of Graph Neural Networks to dynamic graphs. Distributed TGNN training requires efficiently tackling temporal dependency, which often leads to excessive cross-device communication that generates significant redundant data. However, existing systems are unable to remove the redundancy in data reuse and transfer, and suffer from severe communication overhead in a distributed setting. This paper presents Sven, an algorithm and system co-designed TGNN training library for the end-to-end performance optimization on multi-node multi-GPU systems. Exploiting dependency patterns of TGNN models and characteristics of dynamic graph datasets, we design redundancy-free data organization and load-balancing partitioning strategies that mitigate the redundant data communication and evenly partition dynamic graphs at the vertex level. Furthermore, we develop a hierarchical pipeline mechanism integrating data prefetching, micro-batch pipelining, and asynchronous pipelining to mitigate the communication overhead. As the first scaling study on the memory-based TGNNs training, experiments conducted on an HPC cluster of 64 GPUs show that Sven can achieve up to 1.7x-3.3x speedup over the state-of-art approaches and a factor of up to 5.26x communication efficiency improvement.

Recently, Mixture-of-Experts (MoE) has become one of the most popular techniques to scale pre-trained models to extraordinarily large sizes. Dynamic activation of experts allows for conditional computation, increasing the number of parameters of neural networks, which is critical for absorbing the vast amounts of knowledge available in many deep learning areas. However, despite the existing system and algorithm optimizations, there are significant challenges to be tackled when it comes to the inefficiencies of communication and memory consumption.In this paper, we present the design and implementation of MPipeMoE, a high-performance library that accelerates MoE training with adaptive and memory-efficient pipeline parallelism. Inspired by that the MoE training procedure can be divided into multiple independent sub-stages, we design adaptive pipeline parallelism with an online algorithm to configure the granularity of the pipelining. Further, we analyze the memory footprint breakdown of MoE training and identify that activations and temporary buffers are the primary contributors to the overall memory footprint. Toward memory efficiency, we propose memory reusing strategies to reduce memory requirements by eliminating memory redundancies, and develop an adaptive selection component to determine the optimal strategy that considers both hardware capacities and model characteristics at runtime. We implement MPipeMoE upon PyTorch and evaluate it with common MoE models in a physical cluster consisting of 8 NVIDIA DGX A100 servers. Compared with the state-of-art approach, MPipeMoE achieves up to 2.8× speedup and reduces memory footprint by up to 47% in training large models.

We introduce the sparse modern Hopfield model as a sparse extension of the modern Hopfield model. Like its dense counterpart, the sparse modern Hopfield model equips a memory-retrieval dynamics whose one-step approximation corresponds to the sparse attention mechanism. Theoretically, our key contribution is a principled derivation of a closed-form sparse Hopfield energy using the convex conjugate of the sparse entropic regularizer. Building upon this, we derive the sparse memory retrieval dynamics from the sparse energy function and show its one-step approximation is equivalent to the sparse-structured attention. Importantly, we provide a sparsity-dependent memory retrieval error bound which is provably tighter than its dense analog. The conditions for the benefits of sparsity to arise are therefore identified and discussed. In addition, we show that the sparse modern Hopfield model maintains the robust theoretical properties of its dense counterpart, including rapid fixed point convergence and exponential memory capacity. Empirically, we use both synthetic and real-world datasets to demonstrate that the sparse Hopfield model outperforms its dense counterpart in many situations.

With the large-scale growth of medical data, Chinese medical named entity recognition (NER) has attracted extensive attention, which plays a crucial prerequisite role in extracting high-quality medical information. Many studies have confirmed that additional lexical and radical information can improve the Chinese medical NER model effectively. Still, a method that can effectively utilize both types of information has been lacking. In this paper, various forms of introducing lexical and radical information into the NER task are deeply investigated. On this basis, TsERL: Two-stage Enhancement of Radical and Lexicon is proposed for Chinese medical NER, which constructs the relation graph among radicals, characters, and words. In terms of lexical enhancement, TsERL utilizes word graph to match characters with dynamic word counts from three aspects according to the match position. TsERL follows adaptive embedding (AE) paradigm that is transferable, while maintaining better performance than dynamic architecture (DA) paradigm. In terms of radical enhancement, TsERL constructs a global radical graph according to the role of radicals in the Chinese system. On the graph, character nodes can perceive the information of other character nodes and radical nodes through multi-hop relationships. With the power of GAT, TsERL can obtain character type information and lexical boundary information from the radical graph and the word graph, respectively. Experiments on two public datasets show that TsERL outperforms other models in performance.

In a Markov decision process (MDP), an agent interacts with the environment via perceptions and actions. During this process, the agent aims to maximize its own gain. Hence, appropriate regulations are often required, if we hope to take the external costs/benefits of its actions into consideration. In this paper, we study how to regulate such an agent by redesigning model parameters that can affect the rewards and/or the transition kernels. We formulate this problem as a bilevel program, in which the lower-level MDP is regulated by the upper-level model designer. To solve the resulting problem, we develop a scheme that allows the designer to iteratively predict the agent’s reaction by solving the MDP and then adaptively update model parameters based on the predicted reaction. The algorithm is first theoretically analyzed and then empirically tested on several MDP models arising in economics and robotics.

As large-scale data analytic becomes norm in various industries, using MapReduce frameworks to analyze ever-increasing volumes of data will keep growing. In turn, this trend drives up the intention to move MapReduce into multi-tenant clouds. However, the application performance of MapReduce can be significantly affected by the time-varying network bandwidth in a shared cluster. Although many recent studies improve MapReduce performance by dynamic scheduling to reduce the shuffle traffic, most of them do not consider the impact by widely existing hierarchical network architectures in data centers. In this article, we propose and design a hierarchical topology (Hit) aware MapReduce scheduler to minimize overall data traffic cost and hence to reduce job execution time. We first formulate the problem as a Topology Aware Assignment (TAA) optimization problem while considering dynamic computing and communication resources in the cloud with hierarchical network architecture. We further develop a synergistic strategy to solve the TAA problem by using the stable matching theory, which ensures the preference of both individual tasks and hosting machines. Finally, we implement the proposed scheduler as a pluggable module on Hadoop YARN and evaluate its performance by testbed experiments and simulations. The testbed experimental results show that Hit-scheduler can improve job completion time by 28 and 11 percent compared to Capacity Scheduler and Probabilistic Network-Aware scheduler, respectively. Our simulations further demonstrate that Hit-scheduler can reduce the traffic cost by 38 percent at most and the average shuffle flow traffic time by 32 percent compared to capacity scheduler. In this article, we have extended Hit-scheduler to a decentralized heuristic scheme to perform the policy-aware allocation in data center environments. Many existing centralized approximation approaches are too complex and infeasible to implement over a data center, which typically include large amounts of servers, containers, switches, and traffic flows. In the extension, we have designed a decentralized heuristic scheme to perform the Policy-Aware Task (PAT) allocation by using existing centralize algorithm to approximately maximize the total gained utility. Finally, the simulation based experimental results show that the proposed PAT policy reduces the communication cost by 33.6 percent compared with the default scheduler in data centers.

Memory resource is a critical bottleneck for large-scale Deep Neural Network (DNN) applications. Hybrid Memory System (HMS) provides a promising solution to increase memory capacity in an affordable way. However, to release the powerful performance of HMS, data migration plays an important role. A typical DNN application has a couple of execution layers, and each requires distinct data objects. Deploying DNN on HMS imposes enormous challenges on data migration strategy and inspires us to pursue smart solutions. To tackle the data migration problem on HMS for DNN applications, we propose a runtime system for HMS that automatically optimizes scalable data migration and exploits domain knowledge on DNN to decide data migrations between the fast and slow memories in HMS. To achieve a better performance in data migrations for DNN training, we introduce a reference distance and location based data management strategy (ReDL) that treats short-lived and long-lived data objects with Idle and Dynamic migration methods, respectively. Using ReDL, DNN training on HMS with a smaller fast memory size can achieve similar performance to the fast memory-only system. The experimental results demonstrate that with configured the size of fast memory to be 20% of each workload’s peak memory consumption, our work achieves a similar performance (at most 9.6% performance difference) to the fast memory-only system. It further achieves an average of 19% and 11% improvement in data locality against the state-of-the-art solutions.

Many deep learning applications deployed in dynamic environments change over time, in which the training models are supposed to be continuously updated with streaming data to guarantee better descriptions of data trends. However, most state-of-the-art learning frameworks support well inofflinetraining methods while omittingonline model updatingstrategies. In this work, we propose and implementiDlaLayer, a thin middleware layer on top of existing training frameworks that streamlines the support and implementation of online deep learning applications. In pursuit of good model quality and fast data incorporation, we design a Data Life Aware model updating strategy (DLA), which builds training data samples according to contributions of data from different life stages, and considers the training cost consumed in model updating. We evaluate iDlaLayer's performance through simulations and experiments based on TensorflowOnSpark with three representative online learning workloads. Our experimental results demonstrate that iDlaLayer reduces the overall elapsed time of ResNet, DeepFM and PageRank by 11.3, 28.2, and 15.2 percent compared to the periodic update strategy, respectively. It further achieves an average 20 percent decrease in training cost and brings about a 5 percent improvement in model quality against the traditional continuous training method.

Memory caches are being widely adopted in today's data-intensive computing frameworks to maximize the benefit of in-memory access. Various memory architectures, designs and management algorithms have been well explored so that application performance can witness speedup by orders of magnitude. For example, Apache Spark provides intermediate data consistency in memory between computation tasks, eliminating a significant amount of disk I/Os and dramatically reducing data processing times. However, the memory space of individual executors is isolated so far in vanilla Spark and apparently inefficient given the fact that these memory demands of different executors vary a lot over tasks, jobs and applications. In this work, we propose a new shared in-Memory cache layer, i.e., iMlayer, among these parallel executors, which are co-hosted on the same slave machine in Apache Spark. It aims to improve the overall hit rate of data blocks by caching and evicting these blocks uniformly across multiple executors. The critical insight of iMlayer is to develop a novel eviction strategy to efficiently manage the shared cache space among executors to maximize the cache hit rate as well as application performance. We evaluate iMlayer based on the three representative workloads from HiBench. Our results demonstrate that iMlayer with the new eviction strategy improves the cache hit rate by 45%, 16% and 27%, effectively reducing the overall job runtime 47%, 43% and 38% compared to vanilla Spark, respectively.

Many deep learning applications deployed in dynamic environments change over time, in which the training models are supposed to be continuously updated with streaming data in order to guarantee better descriptions on data trends. However, most of the state-of-the-art learning frameworks support well in offline training methods while omitting online model updating strategies. In this work, we propose and implement iDlaLayer, a thin middleware layer on top of existing training frameworks that streamlines the support and implementation of online deep learning applications. In pursuit of good model quality as well as fast data incorporation, we design a Data Life Aware model updating strategy (DLA), which builds training data samples according to contributions of data from different life stages, and considers the training cost consumed in model updating. We evaluate iDlaLayer's performance through both simulations and experiments based on TensorflowOnSpark with three representative online learning workloads. Our experimental results demonstrate that iDlaLayer reduces the overall elapsed time of MNIST, Criteo and PageRank by 11.3%, 28.2% and 15.2% compared to the periodic update strategy, respectively. It further achieves an average 20% decrease in training cost and brings about 5 % improvement in model quality against the traditional continuous training method.

Deep neural networks (DNNs) have been widely applied in the field of artificial intelligence, e.g., natural language processing, computer vision, etc. Researchers and industry practitioners typically use GPU to train complex hundred-layers deep networks. However, as the networks going wider and deeper, the limited GPU memory becomes a significant bottleneck, restricting the size of networks to be trained. In the training of DNNs, the intermediate layer outputs are the major contributors to the memory footprint. Offloading and prefetching feature maps is one of the crucial techniques to overcome the GPU memory shortage by utilizing the CPU DRAM as an external buffer for the GPU. However, we find that the layer-by-layer asynchronous approach cannot be effectively applied to the overlap between communication and computation, particularly for nonlinear networks. Furthermore, the default memory management policy could cause high GPU memory fragmentation for the networks with complex nonlinearities. Based on these observations, we adopt an efficient graph analysis and exploit the layered dependency structures to improve the overlap ratio. To achieve minimal memory fragmentation, we design a Group Tensors By Mobility (GTBM) placement policy to allocate tensors on the proposed unified memory pool for data structures with varied data sizes and dynamic dependencies. We implement and evaluate our system, Dymem, on several linear and nonlinear networks. Compared with vDNN and SuperNeurons, our proposed approach can achieve memory cost reduction by up to 31%. The dependency-aware strategy can improve the end-to-end throughput for nonlinear networks by up to 42%.

Decentralized algorithms, e.g., AllReduce, have been widely applied as the synchronization strategy for data-parallel distributed deep learning due to its superior performance over centralized ones. The synchronous Stochastic Gradient Descent (SGD) approach guarantees accuracy for various deep learning models, but its performance suffers from stragglers, i.e., "long-tail effects." The straggler can be caused by the inherent load imbalance from workloads or system heterogeneity. Despite existing optimizations to support centralized algorithms against stragglers, little effort has been explored in decentralized training algorithms.

In memory data analytic frameworks like Apache Spark are employed by an increasing number of diverse applications-such as machine learning, graph computation, and scientific computing, which benefit from the long-running process (e.g. executor) programming model to avoid system I/O overhead. However, existing resource allocation strategies mainly rely on the peak demand normally specified by users. Since the resource usages of long-running applications like iterative computation vary significantly over time, we find that peak-demand-based resource allocation policies lead to low cloud utilization in production environments. In this paper, we present an elastic utilization aware executor provisioning approach for iterative workloads on Apache Spark (i.e., iSpark). It can identify the causes of resource underutilization due to an inflexible resource policy, and elastically adjusts the allocated executors over time according to the real-time resource usage. In general, iterative applications require more computation resources at the beginning stage and their demands for resources diminish as more iterations are completed. iSpark aims to timely scale up or scale down the number of executors in order to fully utilize the allocated resources while taking the dominant factor into consideration. It further preempts the underutilized executors and preserves the cached intermediate data to ensure the data consistency. Testbed evaluations show that iSpark averagely improves the resource utilization of individual executors by 35.2 % compared to vanilla Spark. At the same time, it increases the cluster utilization from 32.1% to 51.3% and effectively reduces the overall job completion time by 20.8% for a set of representative iterative applications.

As MapReduce is becoming increasingly popular in large-scale data analysis, there is a growing need for moving MapReduce into multi-tenant clouds. However, there is an important challenge that the performance of MapReduce applications can be significantly influenced by the time-varying network bandwidth in a shared cluster. Although a few recent studies improve MapReduce performance by dynamic scheduling to reduce the shuffle traffic, most of them do not consider the impact by widely existing hierarchical network architectures in data centers. In this paper, we propose and design a Hierarchical topology (Hit) aware MapReduce scheduler to minimize overall data traffic cost and hence to reduce job execution time. We first formulate the problem as a Topology Aware Assignment (TAA) optimization problem while considering dynamic computing and communication resources in the cloud with hierarchical network architecture. We further develop a synergistic strategy to solve the TAA problem by using the stable matching theory, which ensures the preference of both individual tasks and hosting machines. Finally, we implement the proposed scheduler as a pluggable module on Hadoop YARN and evaluate its performance by testbed experiments and simulations. The experimental results show Hit-scheduler can improve job completion time by 28% and 11% compared to Capacity Scheduler and Probabilistic Network-Aware scheduler, respectively. Our simulations further demonstrate that Hit-scheduler can gain the traffic cost by 38% at most and improve the average shuffle flow traffic time by 32% compared to Capacity scheduler.