Knowledge distillation (KD) has proven highly effective for compressing large models and enhancing the performance of smaller ones. However, its effectiveness diminishes in cross-modal scenarios, such as vision-to-language distillation, where inconsistencies in representation across modalities lead to difficult knowledge transfer. To address this challenge, we propose frequency-decoupled cross-modal knowledge distillation, a method designed to decouple and balance knowledge transfer across modalities by leveraging frequency-domain features. We observed that low-frequency features exhibit high consistency across different modalities, whereas high-frequency features demonstrate extremely low cross-modal similarity. Accordingly, we apply distinct losses to these features: enforcing strong alignment in the low-frequency domain and introducing relaxed alignment for high-frequency features. We also propose a scale consistency loss to address distributional shifts between modalities, and employ a shared classifier to unify feature spaces. Extensive experiments across multiple benchmark datasets show our method substantially outperforms traditional KD and state-of-the-art cross-modal KD approaches.

SRAM is widely used in computing-in-memory (CIM) neural network accelerators because of its relatively mature technology and good compatibility with complementary metal oxide semiconductor logic process. Digital SRAM-CIM is favored by researchers because of its stability and accuracy. However, the current digital SRAM-CIM macro only supports the weight-stationary dataflows, which means the repeated movement of graph data. Some special deep neural network layers, such as depth-wise, make the utilization of computing resources inside CIM low. To overcome these problems, we propose C-CIM, which can switch between input-stationary and weight-stationary dataflows and support matrix multiplication as well as convolution operations with multiple mainstream convolution kernel sizes (1×1, 3×3, 5×5 and 7×7). The C-CIM achieves an average performance of 27.31TOPS/W@8b at a frequency of 1GHz. Experimental results show that our proposed SRAM-CIM successfully outperforms baseline in terms of performance optimization, achieving up to 7.6× performance speedup and up to 86.84% reduction in activation relocation.

Pareto Front Learning (PFL) has been one of the effective means to resolve multi-objective optimization problems through exploring all optimal solutions to learn the entire Pareto front. Pareto Hypernetwork (PHN) is a new promising way to generate the sequence of Pareto-optimal solutions that can be further used as potential solutions to constitute the Pareto front. However, the existing PHN-based approaches suffer from two performance issues: They take as inputs human-crafted preference vector or chunk embedding, rather than the input data samples, and thus vulnerable to data distribution shifts. Such approaches cannot optimize all potential solutions when forming the Pareto front, as they merely optimize the loss pertaining to one single input at a time of optimization round. To improve the quality of the Pareto front, we propose IOP, a novel Idempotent-like Optimization method to learn the entire Pareto front accurately and enhance Hypernetwork's adaptability to distribution shifts. In particular, IOP performs idempotent-like optimization by exploiting manifold space mapping, so that the target networks generated by the optimized Hypernetwork can effectively handle samples with similar distributions of the input samples, without the pre-defined human-crafted inputs. IOP maximizes the Hypervolume indicator that is composed of all potential solutions at a higher level. Experimental results demonstrate that IOP outperforms the state-of-the-art methods by 4.7% on average in producing the Pareto front and has a 10.5% improvement in adaptability.

Recent advances in large language models (LLMs) have intensified the need for serving LLMs that are cost-efficient and QoS-guaranteed. Existing frameworks often co-locate computationally distinct prefill and decode instances on homogeneous GPUs, overlooking their unique resource demands and under-utilizing heterogeneous GPUs. This leads to suboptimal resource utilization and increased capital expenditure. We present Cauchy, a LLM serving framework that adaptively deploys prefill and decode computation to the most suitable heterogeneous GPUs and dynamically schedules user requests. At the core of Cauchy is choosing proper GPU Combo, a conceptual GPU combination encompassing diverse GPU configurations, for their cost efficiency in running prefill-decode pairs. Cauchy deploys a set of combos to satisfy QoS requirements (e.g., goodput) of LLM inference. Cauchy further employs hierarchical scheduling to handle user requests, using opportunistic scheduling within the allocated GPU Combos and a goodput-weighted round-robin policy across GPU Combos. Dynamic autoscaling is used to stabilize the cost-efficiency in the face of surging requests. Experiments show that Cauchy achieves up to a 38.3% improvement in Tokens/USD efficiency over the state-of-the-art baselines, while maintaining strict Service Level Objectives (SLOs). Our work highlights the importance of leveraging workload and GPU heterogeneity to achieve superior cost-efficient LLM serving.

The growing scale and heterogeneity of GPU clusters pose new challenges to deep learning (DL) job scheduling. While existing schedulers primarily focus on GPU utilization, they often ignore multi-dimensional resource demands of DL workloads and lack precise execution time estimation for co-located jobs. While Muri pioneered the use of interleaving execution to improve resource efficiency, it simplified interference when jobs using one resource simultaneously and is agnostic to the deadline constraints. The job grouping also comes to suboptimal when heterogeneous GPU devices are taken into account. In this paper, we propose Cuckoo, a scheduling system that packs deep learning jobs with stringent deadline requirements over a set of heterogeneous GPU devices where multi-dimensional resources are interleaved and shared by a group of jobs. Specifically, the interleaving execution of simultaneous jobs is characterized and modeled through stage-grained execution time estimation considering the runtime performance interference and the impact of GPU heterogeneity on the job performance. The job packing is formulated as a multi-objective optimization problem which is then solved by the maximum weight matching algorithm. Cuckoo then allocates heterogeneous resources to the packed job groups through a graph-based maximum flow and minimum cut algorithm. Experiments show that Cuckoo improves deadline satisfaction rate by 2.38x and reduces average job completion time (JCT) by 1.81x compared with the state-of-the-art approaches. Cuckoo is implemented based on Kubernetes and has been deployed in Kuaishou to serve thousands of model training jobs that can be interleaved on shared heterogeneous GPU clusters

The feature extractor component in today's ML-based traffic analysis applications is becoming a key bottleneck. While mainstream software-based approaches can support flexible feature extraction, they fail to scale to multi-100Gbps network speed easily. Meanwhile, hardware-accelerated solutions can scale to high throughput, but cannot flexibly support generic traffic analysis applications. In this paper, we propose SuperFE, a feature extraction framework that allows users to extract traffic features efficiently and flexibly. SuperFE leverages the capabilities of both new-generation programmable switches and SmartNICs, with three key designs. First, SuperFE presents a user-friendly and extensible interface to support customized feature extraction policies, shielding underlying hardware implementation details and complexities. Second, SuperFE introduces a high-performance multi-granularity key-vector cache system in the programmable switches to batch necessary feature metadata for massive amounts of packets. Third, SuperFE exploits the multi-core parallel and hierarchical memory of SoC-based SmartNICs to achieve efficient feature computation with diverse streaming algorithms. Evaluations using our prototype demonstrate that SuperFE enables various state-of-the-art traffic analysis applications to efficiently extract features from multi-100Gbps raw traffic without compromising detection accuracy, and achieves nearly two orders of magnitude higher throughput than the software-based counterparts.

Digital Compute-in-Memory (CIM) macros revolutionize the Von Neumann architecture by significantly reducing data movements between CPU and memory. However, when dealing with lightweight CNNs with various convolution types, existing GEMM (general matrix multiplication)-oriented solutions suffer from underutilization and large activation traffic, leading to unsatisfied energy and area efficiencies. In this context, we propose CAMO, in which the key contributions are: (1) A novel convolution mapping mechanism suitable for CIM macros, that maximizes data reuse and reduces activation traffic. (2) A convolution-capable CIM macro, that also supports small-scale GEMM. (3) A CIM-based architecture that supports multiple computing modes. The experimental results show that CAMO achieves up to 31.49× performance speedup and 77.5% activation traffic reduction compared to the baseline architecture.

Traditional navigation systems often overlook dynamic social contexts, affecting users' sense of safety and comfort in unfamiliar settings. This challenge requires integrating personalized real-time social data into route planning on resource-constrained mobile and wearable platforms. This paper introduces AuraNav, enhancing navigation by incorporating real-time recognition and individualized social patterns. AuraNav includes two main modules: (i) Social Topology Enhanced Navigation uses real-time familiar individual detection to create dynamic safety corridors by adjusting path costs with an influence field-based metric in A* search; and (ii) Personalized Path Memory mines long-term familiarity data to build heatmaps, offering route recommendations aligned with user habits and comfort zones. This system uses a modular edge-cloud architecture for low-latency sensor data processing and scalable analysis of social-spatial information. The evaluation shows that AuraNav improves routing safety by 17-20% and reduces travel time by 5%, maintaining immediate responsiveness with negligible latency. AuraNav offers a framework for socially aware navigation systems that evolve into personalized and context-aware systems.

The resilience of large-scale AI training platforms are fundamental to enabling contemporary AI innovation and business development. However, with the rapid increase in the scale and complexity of AI model training tasks, anomalies become the norm rather than the exception at scale. Failing to handle them properly may lead to enormous resource waste and prolonged development cycles. Traditional anomaly detection methods struggle to tackle the complex temporal characteristics and extreme class imbalance inherently manifesting in training tasks, and fall short in automated solution to root cause analysis and the follow-up remediation. This paper proposes KAIOPS, an end-to-end automated platform solution for handling anomalies and engineering experience of daily operational maintenance for large-scale AI training clusters at Kuaishou. KAIOPS employs a Temporal Context Encoding mechanism to precisely capture and encode long-term trends and critical temporal context information within fault evolution. The detection model elaborates a dynamic class-weighted loss function for enhancing the detection performance. To deliver a complete end-to-end intelligent processing pipeline, KAIOPS further leverages knowledge graph and LLMs for automated root cause analysis and actionable solution generation. Extensive experiments, on the basis of data collected from Kuaishou’s production-grade training clusters, show the superior performance of our proposed approach. KAIOPS has been deployed in Kuaishou, in both testbed and production grade environments, consisting of with over 10,000 GPUs, and accelerate the reliability assurance for industry-scale model training and serving.

The distributed deep learning training process within large-scale clusters serves as the foundation of contemporary artificial intelligence. However, its inherent characteristics make it particularly sensitive to stragglers, specifically the presence of slow workers, which can significantly decelerate the entire procedure. Observability tools are essential for identifying stragglers within systems. However, the prevailing system profiling tools are either designed for single-node analysis, lacking visibility across multiple workers, or they recognize stragglers but only deliver high-level symptoms, providing engineers with insufficient insight into the underlying causes.We design Kair, a robust production-standard observability tool. Kair uses an innovative hierarchical approach, transitioning from statistical anomaly detection to causal inference. It employs Kolmogorov-Smirnov statistics for the identification of statistically anomalous workers and implements a causal path tracing algorithm to accurately determine the specific operations, such as computation or communication, that are responsible for the delay. Kair has been evaluated in a production cluster of 2,048 NVIDIA A800 GPUs and demonstrated high effectiveness in detecting latent stragglers at the framework level that are often overlooked by conventional tools. It offers precise suggestions that markedly reduce processing inefficiencies and engineering workload.

Serverless computing has gained widespread popularity among developers due to its low cost, fine-grained deployment, and management-free operation. However, when deploying serverless applications in practice, a single function is often insufficient to fulfill complete application requirements. This has led to the emergence of serverless workflows, which orchestrate a series of related serverless functions according to predefined logic. Through our investigation of existing serverless workflow platforms, we identify two major security limitations. First, current serverless workflows cannot guarantee execution integrity—they are unable to detect changes in the function execution order and lack mechanisms to defend against workflow-targeted denial-of-service (DoS) attacks. Second, identity management is typically coarse-grained, often resulting in over-privileged access and lacking support for function-level access control.To address these issues, we propose and implement the LASEFlow, a label-aware security enhancement framework for serverless workflows. A sequential function execution chain is designed based on the Platform Configuration Register (PCR) technique to guarantee the integrity of the execution of workflow in LASEFlow. In addition, a fine-grained function-level access control mechanism is designed to prevent privilege abuse in work-flows. The evaluation demonstrates the effectiveness of LASEFlow against workflow attacks, with an overhead of less than 4% in performance.

The spatio-temporal (ST) mobility patterns derived from trajectory data are crucial for applications such as location-based services and urban analytics. However, releasing these mobility features raises significant privacy concerns, as they may expose sensitive personal location information. Differential privacy (DP) is widely used to safeguard individual privacy during data releases, but existing methods for releasing ST features often suffer from utility loss because their high dimensionality requires injecting substantial noise to meet privacy guarantees. Several recent approaches attempt to address this issue by reducing noise in differentially private spatio-temporal (DPST) features, but they either discard valuable information while compressing noisy data representations or rely solely on restrictive road network topology constraints, resulting in only modest utility improvements. In this paper, we present DPDeno, a post-processing framework designed to significantly enhance the utility of DPST features. First, DPDeno generates synthetic trajectory datasets using public information (e.g., road network data) and applies existing DP methods to create paired DPST (noisy) and ST (clean) features. It then trains a spatio-temporal graph autoencoder (STGAE), which models each feature as a graph, with road segments as nodes and transitions over time as edges. By minimizing node- and edge-level reconstruction losses between the noisy and clean pairs, STGAE learns to refine DPST inputs toward the structural consistency of their clean counterparts, thereby improving their practical utility. The trained model is then used to post-process real DPST features. Importantly, DPDeno preserves the original DP guarantee, as STGAE is trained solely on synthetic data generated from public sources without accessing any private information. Experimental results on two real-world trajectory datasets show that DPDeno significantly improves both the statistical accuracy and practical usability of released mobility features.

Large-scale GPU clusters have been widely used for effectively training both online and offline deep learning (DL) jobs. However, elastic scheduling in most cases of resource schedulers is dedicated for offline model training where resource adjustment is planned ahead of time. The native autoscaling policy is on the basis of pre-defined threshold and, if applied directly in online model training, often suffers from belated resource adjustment, leading to diminished model accuracy. In this paper, we present Kale, a novel elastic GPU scheduling system to improve the performance of online DL model training. Through traffic forecasting and resource-throughput modeling, Kale automatically pinpoints the number of required GPUs that best accommodate the on-the-fly data samples before performing stabilized autoscaling. An advanced data shuffling strategy is further employed for balancing uneven samples among different training workers, thereby improving the runtime efficacy. Experiments show that Kale substantially outperforms the state-of-the-art solutions. Compared with the default HPA autoscaling strategy, Kale reduces the accumulated lag and downtime by 69.2% and 33.1%, respectively, whilst lowering the SLO violation rate from 19.57% to just 2.6%. Kale has been deployed at Kuaishou's production-level GPU clusters and successfully underpins real-time video recommendation and advertisement at scale.

Local differential privacy (LDP), which enables an untrusted server to collect aggregated statistics from distributed users while protecting the privacy of those users, has been widely deployed in practice. However, LDP protocols for frequency estimation are vulnerable to poisoning attacks, in which an attacker can poison the aggregated frequencies by manipulating the data sent from malicious users. Therefore, it is an open challenge to recover the accurate aggregated frequencies from poisoned ones. In this work, we propose LDPRecover, a method that can recover accurate aggregated frequencies from poisoning attacks, even if the server does not learn the details of the attacks. In LDPRecover, we establish a genuine frequency estimator that theoretically guides the server to recover the frequencies aggregated from genuine users' data by eliminating the impact of malicious users' data in poisoned frequencies. Since the server has no idea of the attacks, we propose an adaptive attack to unify existing attacks and learn the statistics of the malicious data within this adaptive attack by exploiting the properties of LDP protocols. By taking the estimator and the learning statistics as constraints, we formulate the problem of recovering aggregated frequencies to approach the genuine ones as a constraint inference (CI) problem. Consequently, the server can obtain accurate aggregated frequencies by solving this problem optimally. Moreover, LDPRecover can serve as a frequency recovery paradigm that recovers more accurate aggregated frequencies by integrating attack details as new constraints in the CI problem. Our evaluation on two real-world datasets, three LDP protocols, and untargeted and targeted poisoning attacks shows that LDPRecover is both accurate and widely applicable against various poisoning attacks.

Deep learning recommendation models (DLRM) exploit user behaviors such as clicks, browse footprints, preferences, etc. for improved personalized experiences. However, in the face of the exponential growth of user data, such models require increasing GPU resources that are unaffordable and insufficient in a computing cluster. To improve GPU utilization and facilitate the advances of GPU scheduling algorithms, we present PrecisionProbe, a non-intrusive monitoring and analysis tool that can run upon Kubernetes and conduct sophisticated analytics of GPU resource utilization without altering the existing training code. PrecisionProbe captures fine-grained GPU metrics at the level of individual model layers and allows for a precise understanding of resource consumption patterns by exploring such detailed metrics. The mechanism is crucial for devising effective GPU scheduling algorithms, particularly tailored for DLRM training jobs dependent upon consumption patterns. Experimental results show that the recommendation models, as opposed to CV and NLP models, utilize less FP32 processing but have higher memory interaction frequencies. These findings indicate the unique resource needs of recommendation systems and necessitate the need of performance analytic using PrecisionProbe.

Microservice architecture has become a prevalent paradigm for constructing scalable and flexible cloud-native applications by leveraging the abundant resources of the cloud. However, the topological complexity of microservices poses significant challenges to resource management frameworks that rely on container orchestration. It is paramount to optimize resource utilization within cloud computing clusters while reducing operational costs for service providers. To this end, we present RESCAPE, a framework designed to effectively predict the resource demands of variable microservice workloads. It is instrumental for downstream optimization tasks, particularly heterogeneous resource scheduling, aiming to enhance resource utilization and efficiency. Experiments based on open-source microservice benchmarks such as DeathStarBench and HPC-AI500 demonstrate an average absolute percentage error (MAPE) of 7.9% when forecasting resource needs for the subsequent timestamp, which indicates an adequate precision for resource estimation of microservices.

Detecting time-series anomalies is extremely intricate due to the rarity of anomalies and imbalanced sample categories, which often result in costly and challenging anomaly labeling. Most of the existing approaches largely depend on assumptions of normality, overlooking labeled abnormal samples. While anomaly assumptions based methods can incorporate prior knowledge of anomalies for data augmentation in training classifiers, the adopted random or coarse-grained augmentation approaches solely focus on pointwise anomalies and lack cutting-edge domain knowledge, making them less likely to achieve better performance. This paper introduces CutAddPaste, a novel anomaly assumption-based approach for detecting time-series anomalies. It primarily employs a data augmentation strategy to generate pseudo anomalies, by exploiting prior knowledge of anomalies as much as possible. At the core of CutAddPaste is cutting patches from random positions in temporal subsequence samples, adding linear trend terms, and pasting them into other samples, so that it can well approximate a variety of anomalies, including point and pattern anomalies. Experiments on standard benchmark datasets demonstrate that our method outperforms the state-of-the-art approaches.

Recent advancements in social bot detection have been driven by the adoption of Graph Neural Networks. The social graph, constructed from social network interactions, contains benign and bot accounts that influence each other. However, previous graph-based detection methods that follow the transductive message-passing paradigm may not fully utilize hidden graph information and are vulnerable to adversarial bot behavior. The indiscriminate message passing between nodes from different categories and communities results in excessively homogeneous node representations, ultimately reducing the effectiveness of social bot detectors. In this paper, we propose \SEBot, a novel multi-view graph-based contrastive learning-enabled social bot detector. In particular, we use structural entropy as an uncertainty metric to optimize the entire graph's structure and subgraph-level granularity, revealing the implicitly existing hierarchical community structure. And we design an encoder to enable message passing beyond the homophily assumption, enhancing robustness to adversarial behaviors of social bots. Finally, we employ multi-view contrastive learning to maximize mutual information between different views and enhance the detection performance through multi-task learning. Experimental results demonstrate that our approach significantly improves the performance of social bot detection compared with SOTA methods.

Android is undergoing unprecedented malicious threats daily, but the existing methods for malware detection often fail to cope with evolving camouflage in malware. To address this issue, we present Hawk, a new malware detection framework for evolutionary Android applications. We model Android entities and behavioral relationships as a heterogeneous information network (HIN), exploiting its rich semantic meta-structures for specifying implicit higher order relationships. An incremental learning model is created to handle the applications that manifest dynamically, without the need for reconstructing the whole HIN and the subsequent embedding model. The model can pinpoint rapidly the proximity between a new application and existing in-sample applications and aggregate their numerical embeddings under various semantics. Our experiments examine more than 80 860 malicious and 100 375 benign applications developed over a period of seven years, showing that Hawk achieves the highest detection accuracy against baselines and takes only 3.5 ms on average to detect an out-of-sample application, with the accelerated training time of 50× faster than the existing approach.

Social bot detection is of paramount importance to the resilience and security of online social platforms. The state-of-the-art detection models are siloed and have largely overlooked a variety of data characteristics from multiple cross-lingual platforms. Meanwhile, the heterogeneity of data distribution and model architecture make it intricate to devise an efficient cross-platform and cross-model detection framework. In this paper, we propose FedACK, a new federated adversarial contrastive knowledge distillation framework for social bot detection. We devise a GAN-based federated knowledge distillation mechanism for efficiently transferring knowledge of data distribution among clients. In particular, a global generator is used to extract the knowledge of global data distribution and distill it into each client’s local model. We leverage local discriminator to enable customized model design and use local generator for data enhancement with hard-to-decide samples. Local training is conducted as multi-stage adversarial and contrastive learning to enable consistent feature spaces among clients and to constrain the optimization direction of local models, reducing the divergences between local and global models. Experiments demonstrate that FedACK outperforms the state-of-the-art approaches in terms of accuracy, communication efficiency, and feature space consistency.

Graph Neural Networks (GNNs) are de facto solutions to structural data learning. However, it is susceptible to low-quality and unreliable structure, which has been a norm rather than an exception in real-world graphs. Existing graph structure learning (GSL) frameworks still lack robustness and interpretability. This paper proposes a general GSL framework, SE-GSL, through structural entropy and the graph hierarchy abstracted in the encoding tree. Particularly, we exploit the one-dimensional structural entropy to maximize embedded information content when auxiliary neighbourhood attributes is fused to enhance the original graph. A new scheme of constructing optimal encoding trees are proposed to minimize the uncertainty and noises in the graph whilst assuring proper community partition in hierarchical abstraction. We present a novel sample-based mechanism for restoring the graph structure via node structural entropy distribution. It increases the connectivity among nodes with larger uncertainty in lower-level communities. SE-GSL is compatible with various GNN models and enhances the robustness towards noisy and heterophily structures. Extensive experiments show significant improvements in the effectiveness and robustness of structure learning and node representation learning.

Resource provisioning plays a pivotal role in determining the right amount of infrastructure resource to run applications and reduce the monetary cost. A significant portion of production clusters is now dedicated to long-running applications (LRAs), which are typically in the form of microservices and executed in the order of hours or even months. It is therefore practically important to plan ahead the placement of LRAs in a shared cluster for the minimized number of compute nodes required by them. Existing works on LRA scheduling are often application-agnostic, without particularly addressing the constraining requirements imposed by LRAs, such as co-location affinity constraints and time-varying resource requirements. In this paper, we present an affinity-aware resource provisioning approach for deploying large-scale LRAs in a shared cluster subject to multiple constraints, with the objective of minimizing the number of compute nodes in use. We investigate a broad range of solution algorithms which fall into three main categories: Application-Centric, Node-Centric, and Multi-Node approaches, and tune them for typical large-scale real-world scenarios. Experimental studies driven by the Alibaba Tianchi dataset show that our algorithms can achieve competitive scheduling effectiveness and running time, as compared with the heuristics used by the latest work including Medea and LraSched. Best results are obtained by the Application-Centric algorithms, if the algorithm's running time is of primary concern, and by Multi-Node algorithms, if the solution quality is of primary concern.

In recent years, a surging number of unmanned aerial vehicles (UAVs) are pervasively utilized in many areas. However, the increasing number of UAVs may cause privacy and security issues such as voyeurism and espionage. It is critical for individuals or organizations to manage their behaviors and proactively prevent the misbehaved invasion of unauthorized UAVs through effective anomaly detection. The UAV anomaly detection framework needs to cope with complex signals in the noisy-prone environments and to function with very limited labeled samples. This paper proposesBisSiam, a novel framework that is capable of identifying UAV presence, types and operation modes.BisSiamconverts UAVs signals to bispectrum as the input and exploits a siamese network based contrastive learning model to learn the vector encoding. A sampling mechanism is proposed for optimizing the sample size involved in the model training whilst ensuring the model accuracy without compromising the training efficiency. Finally, we present a similarity-based fingerprint matching mechanism for detecting unseen UAVs without the need of retraining the whole model. Experiment results show that our approach outperforms other baselines and can reach 92.85% accuracy of UAV type detection in unsupervised learning scenarios. 91.4% accuracy can be achieved whenBisSiamis used for detecting the UAV type of the out-of-sample UAVs.

We present LUCE, the first life-long predictive model for automated property valuation. LUCE addresses two critical issues of property valuation: the lack of recent sold prices and the sparsity of house data. It is designed to operate on limited volume of recent house transaction. As a departure from prior work, LUCE organizes the house data in a HIN where graph nodes are house entities and attributes that are important for house price valuation. We employ GCN to extract the spatial information from the HIN, and then use LSTM network to model the temporal dependencies over time. Unlike prior work, LUCE makes effective use of the limited house transactions in the past few months to update valuation information for all house entities. By providing a complete and up-to-date house valuation dataset, LUCE thus massively simplifies the downstream valuation task for the targeting properties. We demonstrate the benefit of LUCE by applying it to large, real-life datasets obtained from the Toronto real estate market. Extensive experimental results show that LUCE not only significantly outperforms prior property valuation methods but also often reaches and sometimes exceeds the valuation accuracy given by independent experts when using the actual realization price as the ground truth.

This article focuses on a simple, yet fundamental question of distributed edge computing: “how to handle IoT traffic with different levels of sensitivity and criticality by satisfying the application-specific latency constraints?” This question arises in the practical deployment of edge computing, where user data can arrive at a much faster rate than that they can be processed by an edge node. Addressing this question is critical for meeting the latency requirement for latency-sensitive applications, but existing approaches are inadequate to the problem. We presentJanus, a multi-level traffic scheduling system for managing multiple data streams with various degrees of latency constraints. At the edge node level,Janususes multi-level queues to manage data streams with different latency constraints. It then allocates the output bandwidth of the edge node according to the requirements of applications in different priority queues, aiming to reduce the queuing and processing delay of latency-sensitive streams while maximizing the edge-node throughput. At the network level,Janusactively redirects incoming data streams to the less-loaded ones to achieve better network-wide load balance and improve the overall throughput. Experiments show thatJanusreduces the latency to only 16.6% of a non-priority based solution and improves the throughput by 1.7x of a state-of-the-art priority-aware data stream scheduling approach.

Social bots are referred to as the automated accounts on social networks that make attempts to behave like humans. While Graph Neural Networks (GNNs) have been massively applied to the field of social bot detection, a huge amount of domain expertise and prior knowledge is heavily engaged in the state-of-the-art approaches to design a dedicated neural network architecture for a specific classification task. Involving oversized nodes and network layers in the model design, however, usually causes the over-smoothing problem and the lack of embedding discrimination. In this article, we propose RoSGAS , a novel R einf o rced and S elf-supervised G NN A rchitecture S earch framework to adaptively pinpoint the most suitable multi-hop neighborhood and the number of layers in the GNN architecture. More specifically, we consider the social bot detection problem as a user-centric subgraph embedding and classification task. We exploit the heterogeneous information network to present the user connectivity by leveraging account metadata, relationships, behavioral features, and content features. RoSGAS uses a multi-agent deep reinforcement learning (RL), 31 pages. mechanism for navigating the search of optimal neighborhood and network layers to learn individually the subgraph embedding for each target user. A nearest neighbor mechanism is developed for accelerating the RL training process, and RoSGAS can learn more discriminative subgraph embedding with the aid of self-supervised learning. Experiments on five Twitter datasets show that RoSGAS outperforms the state-of-the-art approaches in terms of accuracy, training efficiency, and stability and has better generalization when handling unseen samples.

State-of-the-art systolic array-based accelerators adopt the traditional im2col algorithm to accelerate the inference of convolutional layers. However, traditional im2col cannot efficiently support AI backpropagation. Backpropagation in convolutional layers involves performing transposed convolution and dilated convolution, which usually introduces plenty of zero-spaces into the feature map or kernel. The zero-space data reorganization interfere with the continuity of training and incur additional and non-negligible overhead in terms of off- and on-chip storage, access and performance. Since countermeasures for backpropagation are rarely proposed, we propose BP-im2col, a novel im2col algorithm for AI backpropagation, and implement it in RTL on a TPU-like accelerator. Experiments on TPU-like accelerator indicate that BP-im2col reduces the backpropagation runtime by 34.9% on average, and reduces the bandwidth of off-chip memory and on-chip buffers by at least 22.7% and 70.6% respectively, over a baseline accelerator adopting the traditional im2col. It further reduces the additional storage overhead in the backpropagation process by at least 74.78%.

As NN accelerators emerging, many analytical models are presented to help designers to carry out hardware design space exploration. However, these models cannot accurately simulate the systolic array-based NN accelerator due to their pervasiveness or abstraction. In this paper, we propose a compute-centric simulator driven by the execution of events from the tiles of the mapping matrix, which can accurately model the systolic array-based accelerator. The simulator focuses on the conflicts when the tile is used for data access, or various interruptions caused by hardware resource limitations. Experimental results show that the proposed simulator achieves more than 95% accuracy compared to the real scenes.

Deep neural networks (DNNs) with billion-scale parameters have demonstrated impressive performance in solving many tasks. Unfortunately, training a billion-scale DNN is out of the reach of many data scientists because it requires high-performance GPU servers that are too expensive to purchase and maintain. We present STRONGHOLD, a novel approach for enabling large DNN model training with no change to the user code. STRONGHOLD scales up the largest trainable model size by dynamically offloading data to the CPU RAM and enabling the use of secondary storage. It automatically determines the minimum amount of data to be kept in the GPU memory to minimize GPU memory usage. Compared to state-of-the-art offloading-based solutions, STRONGHOLD improves the trainable model size by 1.9x~6. Sx on a 32GB V100 GPU, with 1.2x~3.7x improvement on the training throughput. It has been deployed into production to successfully support large-scale DNN training.

There exists a practical need for incentivizing content providers to cache contents at distributed network edges closer to users. However, this is a particularly challenging problem due to system environments that are uncertain, content placements that couple adjacent time slots, and economic properties that are desired but hard to ensure. In this paper, we present our design of an auction-based incentive mechanism for online edge caching. We formulate the long-term social cost minimization problem as a nonlinear mixed-integer program that addresses bid selections, user request dispatching, content placements, and payment determination in repetitive auctions. To solve this problem online, we devise a greedy approximation algorithm for solving each auction individually, and a lazy-replacement-based online algorithm that ties the series of auctions over time while dynamically pursuing the balance between downloading contents to new cache locations and keeping them at existing locations. We formally prove the approximation ratio for each single auction, the competitive ratio for the long-term social cost, as well as the truthfulness, the individual rationality, and the computational efficiency of our approach. Evaluations with real-world data have also validated and confirmed the practical superiority of our approach over multiple alternative algorithms.

Understanding the interconnected relationships of large-scale information networks like social, scholar and Internet of Things networks is vital for tasks like recommendation and fraud detection. The vast majority of the real-world networks are inherently heterogeneous and dynamic, containing many different types of nodes and edges and can change drastically over time. The dynamicity and heterogeneity make it extremely challenging to reason about the network structure. Unfortunately, existing approaches are inadequate in modeling real-life dynamical networks as they either have strong assumption of a given stochastic process or fail to capture the heterogeneity of network structure, and they all require extensive computational resources. We introduceLime, a better approach for modeling dynamic and heterogeneous information networks.Limeis designed to extract high-quality network representation with significantly lower memory resources and computational time over the state-of-the-arts. Unlike prior work that uses a vector to encode each network node, we exploit the semantic relationships among network nodes to encode multiple nodes with similar semantics in shared vectors. By using many fewer node vectors, our approach significantly reduces the required memory space for encoding large-scale networks. To effectively trade information sharing for reduced memory footprint, we employ the recursive neural network (RsNN) with carefully designed optimization strategies to explore the node semantics in a novel cuboid space. We then go further by showing, for the first time, how an effective incremental learning approach can be developed – with the help of RsNN, our cuboid structure, and a set of novel optimization techniques – to allow a learning framework to quickly and efficiently adapt to a constantly evolving network. We evaluateLimeby applying it to three representative network-based tasks, node classification, node clustering and anomaly detection, performing on three large-scale datasets. We compareLimeagainst eleven prior state-of-the-art approaches for learning network representation. Our extensive experiments demonstrate thatLimenot only reduces the memory footprint by over 80 percent and the processing time over 2x when learning network representation but also delivers comparable performance for downstream processing tasks. We show that our incremental learning method can boost the learning time by up to 20x without compromising the quality of the learned network representation.

Text classification is the most fundamental and essential task in natural language processing. The last decade has seen a surge of research in this area due to the unprecedented success of deep learning. Numerous methods, datasets, and evaluation metrics have been proposed in the literature, raising the need for a comprehensive and updated survey. This paper fills the gap by reviewing the state-of-the-art approaches from 1961 to 2021, focusing on models from traditional models to deep learning. We create a taxonomy for text classification according to the text involved and the models used for feature extraction and classification. We then discuss each of these categories in detail, dealing with both the technical developments and benchmark datasets that support tests of predictions. A comprehensive comparison between different techniques, as well as identifying the pros and cons of various evaluation metrics are also provided in this survey. Finally, we conclude by summarizing key implications, future research directions, and the challenges facing the research area.

Graph Neural Networks (GNNs) have been widely used for the representation learning of various structured graph data, typically through message passing among nodes by aggregating their neighborhood information via different operations. While promising, most existing GNNs oversimplify the complexity and diversity of the edges in the graph and thus are inefficient to cope with ubiquitous heterogeneous graphs, which are typically in the form of multi-relational graph representations. In this article, we propose RioGNN , a novel Reinforced, recursive, and flexible neighborhood selection guided multi-relational Graph Neural Network architecture, to navigate complexity of neural network structures whilst maintaining relation-dependent representations. We first construct a multi-relational graph, according to the practical task, to reflect the heterogeneity of nodes, edges, attributes, and labels. To avoid the embedding over-assimilation among different types of nodes, we employ a label-aware neural similarity measure to ascertain the most similar neighbors based on node attributes. A reinforced relation-aware neighbor selection mechanism is developed to choose the most similar neighbors of a targeting node within a relation before aggregating all neighborhood information from different relations to obtain the eventual node embedding. Particularly, to improve the efficiency of neighbor selecting, we propose a new recursive and scalable reinforcement learning framework with estimable depth and width for different scales of multi-relational graphs. RioGNN can learn more discriminative node embedding with enhanced explainability due to the recognition of individual importance of each relation via the filtering threshold mechanism. Comprehensive experiments on real-world graph data and practical tasks demonstrate the advancements of effectiveness, efficiency, and the model explainability, as opposed to other comparative GNN models.

To accelerate the training of Deep Learning (DL) models, clusters of machines equipped with hardware accelerators such as GPUs are leveraged to reduce execution time. State-of-the-art resource managers are needed to increase GPU utilization and maximize throughput. While co-locating DL jobs on the same GPU has been shown to be effective, this can incur interference causing slowdown. In this article we propose Horus: an interference-aware and prediction-based resource manager for DL systems. Horus proactively predicts GPU utilization of heterogeneous DL jobs extrapolated from the DL model's computation graph features, removing the need for online profiling and isolated reserved GPUs. Through micro-benchmarks and job co-location combinations across heterogeneous GPU hardware, we identify GPU utilization as a general proxy metric to determine good placement decisions, in contrast to current approaches which reserve isolated GPUs to perform online profiling and directly measure GPU utilization for each unique submitted job. Our approach promotes high resource utilization and makespan reduction; via real-world experimentation and large-scale trace driven simulation, we demonstrate that Horus outperforms other DL resource managers by up to 61.5 percent for GPU resource utilization, 23.7-30.7 percent for makespan reduction and 68.3 percent in job wait time reduction.

To achieve a high degree of resource utilization, production clusters need to co-schedule diverse workloads – including both batch analytic jobs with short-lived tasks and long-running applications (LRAs) that execute for a long time frame from hours to months – onto the shared resources. Microservice architecture advances the manifestation of distributed LRAs (DLRAs), comprising multiple interconnected microservices that are executed in long-lived distributed containers and serve massive user requests. Detecting and mitigating QoS violation become even more intractable due to the network uncertainties and latency propagation across dependent microservices. However, current resource managers are only responsible for resource allocation among applications/jobs but agnostic to runtime QoS such as latency at application level. The state-of-the-art QoS-aware scheduling approaches are dedicated for monolithic applications, without considering the temporal-spatio performance variability across distributed microservices. In this paper, we presentToposch, a new scheduling and execution framework to prioritize the QoS of DLRAs whilst balancing the performance of batch jobs and maintaining high cluster utilization through harvesting idle resources.Toposchtracks footprints of every single request across microservices and uses critical path analysis, based on the end-to-end latency graph, to identify microservices that have high risk of QoS violation. Based on microservice and node level risk assessment, we intervene the batch scheduling by adaptively reducing the visible resources to batch tasks and thus delaying their execution to give way to DLRAs. We propose a prediction-based vertical resource auto-scaling mechanism, with the aid of resource-performance modeling and fine-grained resource inference and access control, for prompt recovery of QoS violation. A cost-effective task preemption is leveraged to ensure a low-cost task preemption and resource reclamation during the auto-scaling.Toposchis integrated with Apache YARN and experiments show thatToposchoutperforms other baselines in terms of performance guarantee of DLRAs, at an acceptable cost of batch job slowdown. The tail latency of DLRAs is merely 1.12x of the case of executing alone on average inToposchwith a 26% JCT increase of Spark analytic jobs.

Striking a balance between improved cluster utilization and guaranteed application QoS is a long-standing research problem in cluster resource management. The majority of current solutions require a large number of sandboxed experimentation for different workload combinations and leverage them to predict possible interference for incoming workloads. This results in non-negligible time complexity that severely restricts its applicability to complex workload co-locations. The nature of pure offline profiling may also lead to model aging problem that drastically degrades the model precision. In this paper, we present Perph, a runtime agent on a per node basis, which decouples ML-based performance prediction and resource inference from centralized scheduler. We exploit the sensitivity of long-running applications to multi-resources for establishing a relationship between resource allocation and consequential performance. We use Online Gradient Boost Regression Tree (OGBRT) to enable the continuous model evolution. Once performance degradation is detected, resource inference is conducted to work out a proper slice of resources that will be reallocated to recover the target performance. The integration with Node Manager (NM) of Apache YARN shows that the throughput of Kafka data-streaming application is 2.0x and 1.82x times that of isolation execution schemes in native YARN and pure cgroup cpu subsystem. In TPC-C benchmarking, the throughput can also be improved by 35% and 23% respectively against YARN native and cgroup cpu subsystem.

Events are happening in real world and real time, which can be planned and organized for occasions, such as social gatherings, festival celebrations, influential meetings, or sports activities. Social media platforms generate a lot of real-time text information regarding public events with different topics. However, mining social events is challenging because events typically exhibit heterogeneous texture and metadata are often ambiguous. In this article, we first design a novel event-based meta-schema to characterize the semantic relatedness of social events and then build an event-based heterogeneous information network (HIN) integrating information from external knowledge base. Second, we propose a novel Pairwise Popularity Graph Convolutional Network, named as PP-GCN, based on weighted meta-path instance similarity and textual semantic representation as inputs, to perform fine-grained social event categorization and learn the optimal weights of meta-paths in different tasks. Third, we propose a streaming social event detection and evolution discovery framework for HINs based on meta-path similarity search, historical information about meta-paths, and heterogeneous DBSCAN clustering method. Comprehensive experiments on real-world streaming social text data are conducted to compare various social event detection and evolution discovery algorithms. Experimental results demonstrate that our proposed framework outperforms other alternative social event detection and evolution discovery techniques.

CNNs, RNNs, GCNs, and CapsNets have shown significant insights in representation learning and are widely used in various text mining tasks such as large-scale multi-label text classification. Most existing deep models for multi-label text classification consider either the non-consecutive and long-distance semantics or the sequential semantics. However, how to coherently take them into account is still far from studied. In addition, most existing methods treat output labels as independent medoids, ignoring the hierarchical relationships among them, which leads to a substantial loss of useful semantic information. In this paper, we propose a novel hierarchical taxonomy-aware and attentional graph capsule recurrent CNNs framework for large-scale multi-label text classification. Specifically, we first propose to model each document as a word order preserved graph-of-words and normalize it as a corresponding word matrix representation preserving both the non-consecutive, long-distance and local sequential semantics. Then the word matrix is input to the proposed attentional graph capsule recurrent CNNs for effectively learning the semantic features. To leverage the hierarchical relations among the class labels, we propose a hierarchical taxonomy embedding method to learn their representations, and define a novel weighted margin loss by incorporating the label representation similarity. Extensive evaluations on three datasets show that our model significantly improves the performance of large-scale multi-label text classification by comparing with state-of-the-art approaches.

Balancing resource utilization and application QoS is a long-standing research topic in cluster resource management. Big data YARN clusters need to co-schedule diverse workloads on shared resources including batch processing jobs, streaming jobs, and other long-running applications such as web services, database services, etc. Current resource managers are only responsible for resource allocation among applications/jobs but completely unaware of runtime QoS requirements of interactive and latency-sensitive applications. Prior works to maximize the QoS of monolithic applications ignore inherent dependencies and temporal-spatio performance variability of components, characteristics of distributed applications primarily driven by microservices. In this paper, we present Toposch, a new resource management system to adaptively co-locate batch tasks and microservices by harvesting runtime latency. In particular, Toposch tracks full footprints of every request across microservices over time. A latency graph is periodically generated for identifying victim microservices through an end-to-end latency critical path analysis. We then exploit per-microservice and per-node risk assessment to gauge the visible resources to the capacity scheduler in YARN. Execution of batch tasks are adaptively throttled or delayed, thereby avoiding latency increase due to node over-saturation. TOPOSCH is integrated with YARN and experiments show that the latency of DLRAs can be reduced by up to 39.8% against the default capacity scheduling in YARN.

Recent advances in composing Cloud applications have been driven by deployments of inter-networking heterogeneous microservices across multiple Cloud datacenters. System dependability has been of the upmost importance and criticality to both service vendors and customers. Security, a measurable attribute, is increasingly regarded as the representative example of dependability. Literally, with the increment of microservice types and dynamicity, applications are exposed to aggravated internal security threats and externally environmental uncertainties. Existing work mainly focuses on the QoS-aware composition of native VM-based Cloud application components, while ignoring uncertainties and security risks among interactive and interdependent container-based microservices. Still, orchestrating a set of microservices across datacenters under those constraints remains computationally intractable. This paper describes a new dependable microservice orchestration framework GA-Par to effectively select and deploy microservices whilst reducing the discrepancy between user security requirements and actual service provision. We adopt a hybrid (both whitebox and blackbox based) approach to measure the satisfaction of security requirement and the environmental impact of network QoS on system dependability. Due to the exponential grow of solution space, we develop a parallel Genetic Algorithm framework based on Spark to accelerate the operations for calculating the optimal or near-optimal solution. Large-scale real world datasets are utilized to validate models and orchestration approach. Experiments show that our solution outperforms the greedy-based security aware method with 42.34 percent improvement. GA-Par is roughly 4× faster than a Hadoop-based genetic algorithm solver and the effectiveness can be constantly guaranteed under different application scales.

It is a long-standing challenge to achieve a high degree of resource utilization in cluster scheduling. Resource oversubscription has become a common practice in improving resource utilization and cost reduction. However, current centralized approaches to oversubscription suffer from the issue with resource mismatch and fail to take into account other performance requirements, e.g., tail latency. In this article we present ROSE, a new resource management platform capable of conducting performance-aware resource oversubscription. ROSE allows latency-sensitive long-running applications (LRAs) to co-exist with computation-intensive batch jobs. Instead of waiting for resource allocation to be confirmed by the centralized scheduler, job managers in ROSE can independently request to launch speculative tasks within specific machines according to their suitability for oversubscription. Node agents of those machines can however, avoid any excessive resource oversubscription by means of a mechanism for admission control using multi-resource threshold control and performance-aware resource throttle. Experiments show that in case of mixed co-location of batch jobs and latency-sensitive LRAs, the CPU utilization and the disk utilization can reach 56.34 and 43.49 percent, respectively, but the 95th percentile of read latency in YCSB workloads only increases by 5.4 percent against the case of executing the LRAs alone.

Cluster administrators are facing great pressures to improve cluster utilization through workload co-location. Guaranteeing performance of long-running applications (LRAs), however, is far from settled as unpredictable interference across applications is catastrophic to QoS [2]. Current solutions such as [1] usually employ sandboxed and offline profiling for different workload combinations and leverage them to predict incoming interference. However, the time complexity restricts the applicability to complex co-locations. Hence, this issue entails a new framework to harness runtime performance and mitigate the time cost with machine intelligence: i) It is desirable to explore a quantitative relationship between allocated resource and consequent workload performance, not relying on analyzing interference derived from different workload combinations. The majority of works, however, depend on offline profiling and training which may lead to model aging problem. Moreover, multi-resource dimensions (e.g., LLC contention) that are not completely included by existing works but have impact on performance interference need to be considered [3]. ii) Workload co-location also necessitates fine-grained isolation and access control mechanism. Once performance degradation is detected, dynamic resource adjustment will be enforced and application will be assigned an access to specific slices of each resources. Inferring a "just enough" amount of resource adjustment ensures the application performance can be secured whilst improving cluster utilization.

Increased complexity and scale of virtualized distributed systems has resulted in the manifestation of emergent phenomena substantially affecting overall system performance. This phenomena is known as “Long Tail”, whereby a small proportion of task stragglers significantly impede job completion time. While work focuses on straggler detection and mitigation, there is limited work that empirically studies straggler root-cause and quantifies its impact upon system operation. Such analysis is critical to ascertain in-depth knowledge of straggler occurrence for focusing developmental and research efforts towards solving the Long Tail challenge. This paper provides an empirical analysis of straggler root-cause within virtualized Cloud datacenters; we analyze two large-scale production systems to quantify the frequency and impact stragglers impose, and propose a method for conducting root-cause analysis. Results demonstrate approximately 5 percent of task stragglers impact 50 percent of total jobs for batch processes, and 53 percent of stragglers occur due to high server resource utilization. We leverage these findings to propose a method for extreme straggler detection through a combination of offline execution patterns modeling and online analytic agents to monitor tasks at runtime. Experiments show the approach is capable of detecting stragglers less than 11 percent into their execution lifecycle with 95 percent accuracy for short duration jobs.

A long-standing challenge in cluster scheduling is to achieve a high degree of utilization of heterogeneous resources in a cluster. In practice there exists a substantial disparity between perceived and actual resource utilization. A scheduler might regard a cluster as fully utilized if a large resource request queue is present, but the actual resource utilization of the cluster can be in fact very low. This disparity results in the formation of idle resources, leading to inefficient resource usage and incurring high operational costs and an inability to provision services. In this paper we present a new cluster scheduling system, ROSE, that is based on a multi-layered scheduling architecture with an ability to over-subscribe idle resources to accommodate unfulfilled resource requests. ROSE books idle resources in a speculative manner: instead of waiting for resource allocation to be confirmed by the centralized scheduler, it requests intelligently to launch tasks within machines according to their suitability to oversubscribe resources. A threshold control with timely task rescheduling ensures fully-utilized cluster resources without generating potential task stragglers. Experimental results show that ROSE can almost double the average CPU utilization, from 36.37% to 65.10%, compared with a centralized scheduling scheme, and reduce the workload makespan by 30.11%, with an 8.23% disk utilization improvement over other scheduling strategies.

Current Cloud clusters often consist of heterogeneous machine nodes, which can trigger performance challenges such as the task straggler problem, whereby a small subset of parallel tasks running abnormally slower than the other sibling ones. The straggler problem leads to extended job response and deteriorates system throughput. Poor performance nodes are more likely to engender stragglers, and can undermine straggler mitigation effectiveness. For example, as the dominant mechanism for straggler alleviation, speculative execution functions by creating redundant task replicas on other machine nodes as soon as a straggler is detected. When speculative copies are assigned onto the poor performance nodes, it is hard for them to catch up with the stragglers compared to replicas run on fast nodes. And due to the fact that the performance heterogeneity is caused not only by static attribute variations such as physical capacity, but also dynamic characteristic uctuations such as contention level, analyzing node performance is important yet challenging. In this paper we develop ML-NA, a Machine Learning based Node performance Analyzer. By leveraging historical parallel tasks execution log data, ML-NA classies cluster nodes into different categories and predicts their performance in the near future as a scheduling guide to improve speculation effectiveness and minimize task straggler generation. We consider MapReduce as a representative framework to perform our analysis, and use the published OpenCloud trace as a case study to train and to evaluate our model. Results show that ML-NA can predict node performance categories with an average accuracy up to 92.86%.

Large-scale distributed systems deployed as Cloud datacenters are capable of provisioning service to consumers with diverse business requirements. Providers face pressure to provision uninterrupted reliable services while reducing operational costs due to significant software and hardware failures. A widely adopted means to achieve such a goal is using redundant system components to implement user-transparent failover, yet its effectiveness must be balanced carefully without incurring heavy overhead when deployed—an important practical consideration for complex large-scale systems. Failover techniques developed for Cloud systems often suffer serious limitations, including mandatory restart leading to poor cost-effectiveness, as well as solely focusing on crash failures, omitting other important types, such as timing failures and simultaneous failures. This paper addresses these limitations by presenting a new approach to user-transparent failover for massive-scale systems. The approach uses soft-state inference to achieve rapid failure recovery and avoid unnecessary restart, with minimal system resource overhead. It also copes with different failures, including correlated and simultaneous events. The proposed approach was implemented, deployed and evaluated within Fuxi system, the underlying resource management system used within Alibaba Cloud. Results demonstrate that our approach tolerates complex failure scenarios while incurring at worst 228.5 microsecond instance overhead with 1.71 percent additional CPU usage.

The Resource Description Framework (RDF) andSPARQL query language are gaining increasing popularity andacceptance. The ever-increasing RDF data has reached a billionscale of triples, resulting in the proliferation of distributed RDFstore systems within the Semantic Web community. However, theelasticity and performance issues are still far from settled inface of data volume explosion and workload spike. In addition, providers face great pressures to provision uninterrupted reliablestorage service whilst reducing the operational costs due to avariety of system failures. Therefore, how to efficiently realizesystem fault tolerance remains an intractable problem. In this paper, we introduce ScalaRDF, a distributed and elastic in-memoryRDF triple store to provision a fault-tolerant and scalable RDFstore and query mechanism. Specifically, we describe a consistenthashing protocol that optimizes the RDF data placement, dataoperations (especially for online RDF triple update operations)and achieves an autonomously elastic data re-distribution in theevent of cluster node joining or departing, avoiding the holisticoscillation of data storage. In addition, the data store is ableto realize a rapid and transparent failover through replicationmechanism which stores in-memory data replica in the next hashhop. The experiments demonstrate that query time and updatetime are reduced by 87% and 90% respectively compared to otherapproaches. For an 18G source dataset, the data redistributiontakes at most 60 seconds when system scales out and at most 100seconds for recovery when nodes undergo crash-stop failures.

The reliability of data and services hosted in a virtual machine (VM) is a top concern in cloud computing environment. Continuous snapshots reduces the data loss in case of failures, and thus is prevailing for providing protection for long-running systems. However, existing methods suffer from long VM downtime, long snapshot interval and significant performance overhead. In this paper, we present ConSnap, a system designed to enable taking fine-grained continuous snapshots of virtual machines without compromising VM performance. First, ConSnap adopts the COW (copy-on-write) manner to save the memory pages in a lazy way, and thus decrease the snapshot interval to dozens of milliseconds. Second, we only save the incremental memory pages on the basis of the last snapshot in each epoch to reduce the snapshot duration, and thus mitigate VM performance loss. Third, we propose a multi-granularity space reclamation strategy, which merges the unused snapshot files to achieve storage space saving, as well as fast recovery. We have implemented ConSnap on QEMU/KVM and conducted several experiments to verify its effectiveness. Compared with the stop-and-copy based incremental snapshots, ConSnap reduces the performance loss by 71.1% ~ 10.2% under Compilation workload, and 14.5% ~ 4.7% for the Ftp workload, when the interval varies from 1s to 60s.

Virtualization is one of the most fascinating techniques because it can facilitate the infrastructure management and provide isolated execution for running workloads. Despite the benefits gained from virtualization and resource sharing, improved resource utilization is still far from settled due to the dynamic resource requirements and the widely-used over-provision strategy for guaranteed QoS. Additionally, with the emerging demands for big data analytic, how to effectively manage hybrid workloads such as traditional batch task and long-running virtual machine (VM) service needs to be dealt with. In this paper, we propose a system to combine long-running VM service with typical batch workload like MapReduce. The objectives are to improve the holistic cluster utilization through dynamic resource adjustment mechanism for VM without violating other batch workload executions. Furthermore, VM migration is utilized to ensure high availability and avoid potential performance degradation. The experimental results reveal that the dynamically allocated memory is close to the real usage with only 10% estimation margin, and the performance impact on VM and MapReduce jobs are both within 1%. Additionally, at most 50% increment of resource utilization could be achieved. We believe that these findings are in the right direction to solving workload consolidation issues in hybrid computing environments.

Scalability and fault-tolerance are two fundamental challenges for all distributed computing at Internet scale. Despite many recent advances from both academia and industry, these two problems are still far from settled. In this paper, we present Fuxi, a resource management and job scheduling system that is capable of handling the kind of workload at Alibaba where hundreds of terabytes of data are generated and analyzed everyday to help optimize the company's business operations and user experiences. We employ several novel techniques to enable Fuxi to perform efficient scheduling of hundreds of thousands of concurrent tasks over large clusters with thousands of nodes: 1) an incremental resource management protocol that supports multi-dimensional resource allocation and data locality; 2) user-transparent failure recovery where failures of any Fuxi components will not impact the execution of user jobs; and 3) an effective detection mechanism and a multi-level blacklisting scheme that prevents them from affecting job execution. Our evaluation results demonstrate that 95% and 91% scheduled CPU/memory utilization can be fulfilled under synthetic workloads, and Fuxi is capable of achieving 2.36T-B/minute throughput in GraySort. Additionally, the same Fuxi job only experiences approximately 16% slowdown under a 5% fault-injection rate. The slowdown only grows to 20% when we double the fault-injection rate to 10%. Fuxi has been deployed in our production environment since 2009, and it now manages hundreds of thousands of server nodes.

Co-allocated workloads in a virtualized computing environment often have to compete for resources, thereby suffering from performance interference. While this phenomenon has a direct impact on the Quality of Service provided to customers, it also changes the patterns of resource utilization and reduces the amount of work per Watt consumed. Unfortunately, there has been only limited research into how performance interference affects energy-efficiency of servers in such environments. In reality, there is a highly dynamic and complicated correlation among resource utilization, performance interference and energy-efficiency. This paper presents a comprehensive analysis that quantifies the negative impact of performance interference on the energy-efficiency of virtualized servers. Our analysis methodology takes into account the heterogeneous workload characteristics identified from a real Cloud environment. In particular, we investigate the impact due to different workload type combinations and develop a method for approximating the levels of performance interference and energy-efficiency degradation. The proposed method is based on profiles of pair combinations of existing workload types and the patterns derived from the analysis. Our experimental results reveal a non-linear relationship between the increase in interference and the reduction in energy-efficiency as well as an average precision within +/-5% of error margin for the estimation of both parameters. These findings provide vital information for research into dynamic trade-offs between resource utilization, performance, and energy-efficiency of a data center.

Virtualization is one of the main technologies used for improving resource efficiency in datacenters; it allows the deployment of co-existing computing environments over the same hardware infrastructure. However, the co-existing of environments — along with management inefficiencies — often creates scenarios of high-competition for resources between running workloads, leading to performance degradation. This phenomenon is known as Performance Interference, and introduces a non-negligible overhead that affects both a datacenter's Quality of Service and its energy-efficiency. This paper introduces a novel approach to workload allocation that improves energy-efficiency in Cloud datacenters by taking into account their workload heterogeneity. We analyze the impact of performance interference on energy-efficiency using workload characteristics identified from a real Cloud environment, and develop a model that implements various decision-making techniques intelligently to select the best workload host according to its internal interference level. Our experimental results show reductions in interference by 27.5% and increased energy-efficiency up to 15% in contrast to current mechanisms for workload allocation.