Graph Neural Network (GNN) on streaming graphs has gained increasing popularity. However, its practical deployment remains challenging, as the inference process relies on Runtime Embedding Computation (RTEC) to capture recent graph changes. This process incurs heavyweight multi-hop graph traversal overhead, which significantly undermines computation efficiency. We observe that the intermediate results for large portions of the graph remain unchanged during graph evolution, and thus redundant computations can be effectively eliminated through carefully designed incremental methods. In this work, we propose an efficient framework for incrementalizing RTEC on streaming graphs.The key idea is to decouple GNN computation into a set of generalized, fine-grained operators and safely reorder them, transforming the expensive full-neighbor GNN computation into a more efficient form over the affected subgraph. With this design, our framework preserves the semantics and accuracy of the original full-neighbor computation while supporting a wide range of GNN models with complex message-passing patterns. To further scale to graphs with massive historical results, we develop a GPU-CPU co-processing system that offloads embeddings to CPU memory with communication-optimized scheduling. Experiments across diverse graph sizes and GNN models show that our method reduces computation by 64%-99% and achieves 1.7x-145.8x speedups over existing solutions.

The inherent connectivity and dependency of graph-structured data, combined with its unique topology-driven access patterns, pose fundamental challenges to conventional data replication and request routing strategies in geo-distributed cloud storage systems. In this paper, we propose GeoLayer, a geo-distributed graph storage framework that jointly optimizes graph replica placement and pattern request routing. We first construct a latency-aware layered graph architecture that decomposes the graph topology into multiple layers, aiming to reduce the decision space and computational complexity of the optimization problem, while mitigating the impact of network heterogeneity in geo-distributed environments. Building on the layered graph, we introduce an overlap-centric replica placement scheme to accommodate the diversity of graph pattern accesses, along with a directed heat diffusion model that captures heat conduction and superposition effects to guide data allocation. For request routing, we develop a stepwise layered routing strategy that performs progressive expansion over the layered graph to efficiently retrieve the required data. Experimental results show that, compared to state-of-the-art replica placement and routing schemes, GeoLayer achieves a 1.34x - 3.67x improvement in response times for online graph pattern requests and a 1.28x - 3.56x speedup in offline graph analysis performance.

Various graph models have emerged to meet diverse application needs, each with unique characteristics and specialties. Managing and analyzing graph data inevitably requires interactions across different models to serve upstream business requirements. Therefore, an Extract-Transform-Load (ETL) tool designed to bridge different graph models is desired. In this paper, we propose GETL, a generalized graph ETL framework capable of automatically identifying graph model schemas and performing seamless data conversion among RDF, RDF-star, labeled property graph, and the relational model. This is attributed to GETL's unified graph representation model, constructed as nested label, entity pairs, offering powerful capabilities in graph representation and model compatibility. Additionally, we develop a unified programming interface to support complex graph transformation tasks. It is built upon the Gremlin syntax and provides strong expressive capabilities. Finally, our evaluation demonstrates that GETL outperforms state-of-the-art solutions in terms of model conversion efficiency and data manipulation language (DML) intelligibility.

A great number of graph analysis algorithms involve iterative computations, which dominate the runtime. Accelerating iterative graph computations has become the key to improving the performance of graph algorithms. While numerous studies have focused on reducing the runtime of each iteration to improve efficiency, the optimization of the number of iterations is often overlooked. In this work, we first establish a correlation between vertex processing order and the number of iterations, providing an opportunity to reduce the number of iterations. We propose a metric function to evaluate the effectiveness of vertex processing order in accelerating iterative computations. Leveraging this metric, we propose a novel graph reordering method, GoGraph, which constructs an efficient vertex processing order. Additionally, for evolving graphs, we further propose a metric function designed to evaluate the effectiveness of vertex processing orders in response to graph changes and provide three optional methods for dynamically adjusting the vertex processing order. Our experimental results illustrate that GoGraph sur passes current state-of-the-art reordering algorithms, improving runtime by an average of 1.83× (up to 3.34×). Compared to traditional synchronous computation methods, our approach enhances the speed of iterative computations by up to 6.30×. In dynamic scenarios, incremental GoGraph can reduce end-to-end time by 43% on average (up to 48%).

Geo-distributed consensus protocols provide high availability and resilience for distributed database services. These protocols group nodes by their data centers to leverage the network topology that spans across multiple data centers, thereby reducing costly cross-datacenter communication. However, they still face performance and scalability challenges due to inefficient log replication mechanisms. 1) These protocols rely on the leader node in each group to perform cross-datacenter log replication, creating a single-node performance bottleneck. 2) Byzantine receivers can behave arbitrarily, forcing the group leader to send multiple log copies during replication to prevent loss, thus causing redundant transmissions. 3) Since all groups must execute these logs in the same order, synchronizations across groups are necessary to maintain consistency when multiple groups are proposing concurrently, which also slow down log replication. This paper presents MassBFT, a Byzantine fault-tolerant geo-consensus protocol that achieves high performance and scalability. We design an encoded bijective log replication to eliminate the leader bottleneck and reduce the cross-datacenter network consumption. We also propose asynchronous log ordering to eliminate synchronization across groups. Experimental results show that MassBFT is scalable, fault-tolerant, and outperforms state-of-the-art protocols with 5.49-29.96 times higher throughput under YCSB, SmallBank, and TPC-C workloads.

The parameter size of modern large language models (LLMs) can be scaled up to the trillion-level via the sparsely-activated Mixture-of-Experts (MoE) technique to avoid excessive increase of the computational costs. To further improve training efficiency, pipelining computation and communication has become a promising solution for distributed MoE training. However, existing work primarily focuses on scheduling tasks within the MoE layer, such as expert computing and all-to-all (A2A) communication, while neglecting other key operations including multi-head attention (MHA) computing, gating, and all-reduce communication. In this paper, we propose FlowMoE, a scalable framework for scheduling multi-type task pipelines. First, FlowMoE constructs a unified pipeline to consistently scheduling MHA computing, gating, expert computing, and A2A communication. Second, FlowMoE introduces a tensor chunk-based priority scheduling mechanism to overlap the all-reduce communication with all computing tasks. We implement FlowMoE as an adaptive and generic framework atop PyTorch. Extensive experiments with 675 typical MoE layers and four real-world MoE models across two GPU clusters demonstrate that our proposed FlowMoE framework outperforms state-of-the-art MoE training frameworks, reducing training time by14%-57%, energy consumption by 10%-39%, and memory usage by 7%-32%. FlowMoE’s code is anonymously available at https://anonymous.4open.science/r/FlowMoE.

Existing federated learning frameworks are primarily designed for single-modal data. However, real-world scenarios require processing multi-modal data on heterogeneous devices. The gap between existing methods and real-world scenarios presents challenges in processing multimodal data on heterogeneous devices, significantly impacting model training efficiency. To address these issues, we propose a multimodal federated learning framework, which integrates multimodal algorithms with a semi-synchronous training method. The multimodal algorithm trains local autoencoders on different data modalities. By leveraging the similarity of encodings across different modalities with the same data labels, we further train and aggregate these local autoencoders into a global autoencoder, which is then deployed on the blockchain to perform downstream classification tasks. In the semi-synchronous training method, each device updates its parameters independently during a round. At the end of each round, a global aggregation combines the updates from devices. We conduct an empirical evaluation of our framework on various multimodal datasets, including Opportunity (Opp) Challenge, mHealth, and UR Fall Detection datasets. Experimental results demonstrate that our federated learning framework, outperforms the state-of-the-art multimodal frameworks on three multimodal datasets, achieving an average accuracy improvement of 9.07%. Furthermore, in terms of training speed, MMFed is obviously superior to synchronization strategies when it is extended to a large number of clients.

Dynamic Graph Neural Networks (DGNNs) are effective at capturing multidimensional data and enable many important applications. As model training is computationally intensive, distributed DGNN training is employed to accommodate large data. Also, when training DGNNs, so-called sliding window training is used predominantly, as it enhances both accuracy and efficiency. However, current distributed frameworks-such as snapshot partitioning, chunk-based partitioning, and L -hop cache-based communication-free vertex partitioning-are inherently incompatible with sliding window training. While communication-based vertex partitioning supports sliding window training, its design for static graphs limits the effectiveness in distributed DGNN training. Specifically, existing partitioning strategies fail to optimize communication across snapshots, while existing cache reuse and communication scheduling strategies ignore opportunities for optimization between sliding windows. To support distributed sliding window training, we present SWASH, a scalable and flexible communication framework that utilizes a S liding W indow-based c A che SH aring technique. Specifically, we propose a flexible communication framework that supports ratio adjustment and timing selection, as well as hyperparameter settings and adaptive scheduling. We also propose a lightweight partitioning strategy tailored to sliding window-based DGNN training to reduce both partitioning and communication overheads. Finally, to alleviate decreases in accuracy due to reduced communication, we propose a cache-sharing technique based on sliding windows for sharing boundary vertex embeddings. Comprehensive experiments show that SWASH is capable of training speedups of an average of 9.44× over state-of-the-art frameworks while maintaining the accuracy of fully communicating, non-caching training frameworks.

Graph-based Retrieval-Augmented Generation (GraphRAG) has emerged as a promising paradigm for enhancing LLM reliability by enabling multi-hop reasoning over graph-structured knowledge. However, existing LLMs struggle to efficiently process graph-structured inputs, as traditional attention mechanisms are sequence-based and introduce significant redundancy when serializing graphs into prompt sequences, leading to excessive computation and memory overhead. To address this, we introduce dependency attention, a novel graph-aware attention mechanism that restricts attention computation to token pairs with structural dependencies in the retrieved subgraph. Unlike standard self-attention that computes fully connected interactions, dependency attention prunes irrelevant token pairs and reuses computations along shared relational paths, substantially reducing inference overhead. Building on this idea, we develop DepCache, a KV cache management framework tailored for dependency attention. DepCache enables efficient KV cache reuse through (i) a graph-based KV cache reuse strategy that aligns KV caches across varying prompt contexts, enabling efficient cross-request reuse in GraphRAG, and (ii) a locality-aware replacement policy that leverages spatial and temporal access patterns to improve KV cache hit rate. Evaluations across diverse models and datasets show that DepCache improves LLM inference throughput by 1.5×-5.0× and reduces time-to-first-token latency by up to 3.2×, without compromising generation accuracy.

Graph neural networks (GNNs) have emerged as a promising method for learning from graph data, but large-scale GNN training requires extensive memory and computation resources. To address this, researchers have proposed using multi-GPU processing, which partitions graph data across GPUs for parallel training. However, vertex dependencies in multi-GPU GNN training lead to significant neighbor replications across GPUs, increasing memory consumption. The substantial intermediate data generated during training further exacerbates this issue. Neighbor replication and intermediate data constitute the primary memory consumption in GNN training (i.e., typically accounting for over 80%). In this work, we propose GNN task parallelism for multi-GPU GNN training, which reduces neighbor replication by partitioning training tasks in each layer across different GPUs rather than partitioning the graph structure. This approach only partitions the graph data within individual GPUs, reducing the memory requirements of single tasks while overlapping subgraph computation across different GPUs. Shared neighbor embeddings among different subgraphs can be efficiently reused within a single GPU. Additionally, we employ a task-decoupled GNN training framework, which decouples different training tasks to manage their associated intermediate data independently and release it as early as possible to reduce memory usage. By integrating these techniques, we propose a multi-GPU GNN training system, NeutronTask. Experimental results on a 4×A5000 GPU server show that NeutronTask effectively supports billion-scale full-graph GNN training. For small graphs where the training data fits into the GPUs, NeutronTask achieves 1.27× - 5.47× speedup compared to state-of-the-art GNN systems including NeutronStar and Sancus.

Leveraging GPUs' high parallelism can significantly improve the real-time computation efficiency of streaming graph processing. However, when a large-scale graph exceeds GPU memory capacity, CPU-GPU cooperative processing often results in substantial and irregular CPU-to-GPU data transfer overhead. This stems from the extensive redundant graph accesses during continuous computation, which can hardly be addressed by existing solutions. In this work, we present Grapin, an out-of-memory GPU streaming graph processing system designed to minimize graph data transfer via two effective techniques for eliminating redundant accesses: (1) Extending advanced incremental processing algorithms to GPUs by converting their heavyweight data dependency processing into GPU-friendly forms, eliminating redundant graph accesses from the computation side; and (2) providing a lightweight yet efficient GPU hot subgraph management framework that finely caches the frequently accessed dynamic subgraphs in a vertex-centric manner. Experimental results demonstrate that Grapin can efficiently process large-scale streaming graphs with billions of edges on a single NVIDIA A5000 GPU. Enabling incremental computation reduces data transfer by 61%, and the integration of GPU hot subgraph reuse further reduces the remaining transfer by 72%, resulting in a total reduction of 89%. Compared with CPU-based solutions, Grapin achieves speedups ranging from 1.8x to 96.9x (17.9x on average).

Existing disaggregated databases separate execution and storage layers, enabling independent and elastic scaling of resources. In most cases, this design makes transaction concurrency control (CC) a critical bottleneck, which demands significant computing resources for concurrent conflict management and struggles to scale due to the coordination overhead for concurrent conflict resolution. Coupling CC with execution or storage limits performance and elasticity, as CC's resource needs do not align with the free scaling of the transaction execution layer or the storage-bound data layer. This paper proposes Concurrency Control as a Service (CCaaS), which decouples CC from databases, building an execution-CC-storage three-layer decoupled database, allowing independent scaling and upgrades for improved elasticity, resource utilization, and development agility. However, adding a new layer increases latency due to the shift in communication from hardware to network. To address this, we propose a Sharded Multi-Write OCC (SM-OCC) algorithm with an asynchronous log push-down mechanism to minimize network communications overhead and transaction latency. Additionally, we implement a multi-write architecture with a deterministic conflict resolution method to reduce coordination overhead in the CC layer, thereby improving scalability. CCaaS is designed to be connected by a variety of execution and storage engines. Existing disaggregated databases can be revolutionized with CCaaS to achieve high elasticity, scalability, and high performance. Results show that CCaaS achieves 1.02–3.11X higher throughput and 1.11–2.75X lower latency than SoTA disaggregated databases.

Counting triangles is an important topic in many practical applications, such as anomaly detection, community search, and recommendation systems. For triangle counting in large and dynamic graphs, recent work has focused on distributed streaming algorithms. These works assume that the graph is processed in the same location, while in reality, the graph stream may be generated and processed in datacenters that are geographically distributed. This raises new challenges to existing triangle counting algorithms, due to the multi-level heterogeneities in network bandwidth and communication prices in geo-distributed datacenters. In this article, we propose a cost-aware framework named${\sf GeoTri}$based on the Master-Worker-Aggregator architecture, which takes both the cost and performance objectives into consideration for triangle counting in geo-distributed datacenters. The two core parts of this framework are the cost-aware nodes assignment strategy in master, which is critical to obtain node's position and distribute edges reasonably to reduce the cost (i.e., time cost and monetary cost), and cost-aware neighbor transfer strategy among workers, which further eliminates redundancy in data transfers. Additionally, we conduct extensive experiments on seven real-world graphs, and the results demonstrate that${\sf GeoTri}$significantly lowers both runtime and monetary cost while exhibiting nice accuracy and scalability.

The key performance bottleneck of large-scale graph processing on memory-limited GPUs is the host-GPU graph data transfer. Existing GPU-accelerated graph processing frameworks address this issue by managing the active subgraph transfer at runtime. Some frameworks adopt explicit transfer management approaches based on explicit memory copy with filter or compaction. In contrast, others adopt implicit transfer management approaches based on on-demand accesses with the zero-copy mechanism or unified virtual memory. Having made intensive analysis, we find that as the active vertices evolve, the performance of the two approaches varies in different workloads. Due to heavy redundant data transfers, high CPU compaction overhead, or low bandwidth utilization, adopting a single approach often results in suboptimal performance. Moreover, these methods lack effective cache management methods to address the irregular and sparse memory access pattern of graph processing. In this work, we propose a hybrid transfer management approach that takes the merits of both two transfer approaches at runtime. Moreover, we present an efficient vertex-centric graph caching framework that minimizes CPU-GPU communication by caching frequently accessed graph data at runtime. Based on these techniques, we present HytGraph, a GPU-accelerated graph processing framework, which is empowered by a set of effective task-scheduling optimizations to improve performance. Experiments on real-world and synthetic graphs show that HytGraph achieves average speedups of 2.5 ×, 5.0 ×, and 2.0 × compared to the state-of-the-art GPU-accelerated graph processing systems, Grus, Subway, and EMOGI, respectively.

With the rising proliferation of blockchain systems and applications, choosing the appropriate blockchains to deploy applications is critical to achieving optimal performance. Evaluation frameworks provide a systematic approach to assessing and comparing different blockchain systems, guiding application developers to choose the most suitable one. However, existing evaluation frameworks still have limitations that affect their accuracy. First, most frameworks utilize workloads initially de-signed for traditional databases, which fail to capture the unique characteristics and requirements of blockchain systems. Second, these frameworks fail to generate correct results under heavy workloads due to their imbalanced task processing algorithms. Third, existing frameworks are tailored only for non-sharding blockchain architectures, limiting their ability to evaluate diverse blockchains. This paper introduces Hammer, a general blockchain evaluation framework that addresses the above limitations. It consists of two key components: workload prediction and asynchronous task processing. Workload prediction accurately predicts real-world workload trends by expanding the scope of temporal control sequences, providing a more realistic evaluation of blockchain performance. Asynchronous task processing handles heavy-load situations, enabling accurate evaluation of blockchain performance. Extensive experiments on various blockchains under Smallbank workload empower application developers to make informed decisions about blockchain selection and optimization.

GPUs are being applied widely to batch workloads that benefit from the parallel processing capabilities of GPUs. To enable the processing of concurrent batch-based transactions on GPUs, existing systems build dependency graphs during a pre-execution phase to manage read and write operations. However, as dependency-graph maintenance introduces a sub-stantial overhead, there is a need for more efficient transaction support to exploit the power of GPUs more fully for transaction processing. This paper proposes LTPG, a novel GPU-enabled database system that offers increased versatility and efficiency by eliminating the need for predefined read/write-sets. LTPG employs deterministic optimistic concurrency control to ensure correct transaction execution, thus avoiding the maintenance of dependency graphs. The proposed concurrency control simpli-fies transaction processing workflows and avoids the overhead associated with managing dependency graphs, thus resulting in improved efficiency. LTPG divides a workflow into three stages: execution, conflict detection, and write-back, leveraging the parallelism of GPUs. Moreover, several additional optimization strategies are adopted to improve system performance. Experiments with real-world workloads from two benchmarks verify LTPG can achieve effective improvement in the throughput and latency compared to the leading baselines.

Enhancing the efficiency of iterative computation on graphs has garnered considerable attention in both industry and academia. Nonetheless, the majority of efforts focus on expediting iterative computation by minimizing the running time per iteration step, ignoring the optimization of the number of iteration rounds, which is a crucial aspect of iterative compu-tation. We experimentally verified the correlation between the vertex processing order and the number of iterative rounds, thus making it possible to reduce the number of execution rounds for iterative computation. In this paper, we propose a graph reordering method, GoGraph, which can construct a well-formed vertex processing order effectively reducing the number of iteration rounds and, consequently, accelerating iterative computation. Before delving into GoGraph, a metric function is introduced to quantify the efficiency of vertex processing order in accelerating iterative computation. This metric reflects the quality of the processing order by counting the number of edges whose source precedes the destination. GoGraph employs a divide-and-conquer mindset to establish the vertex processing order by maximizing the value of the metric function. Our experimental results show that GoGraph outperforms current state-of-the-art reordering algorithms by 1.83 x on average (up to 3.34 x) in runtime. Compared with traditional synchronous computation, our method improves the iterative computations up to 6.30 x in runtime.

The growing volume of graph data may exhaust the main memory. It is crucial to design a disk-based graph storage system to ingest updates and analyze graphs efficiently. However, existing dynamic graph storage systems suffer from read or write amplification and face the challenge of optimizing both read and write performance simultaneously. To address this challenge, we propose LSMGraph, a novel dynamic graph storage system that combines the write-friendly LSM-tree and the read-friendly CSR. It leverages the multi-level structure of LSM-trees to optimize write performance while utilizing the compact CSR structures embedded in the LSM-trees to boost read performance. LSMGraph uses a new memory structure, MemGraph, to efficiently cache graph updates and uses a multi-level index to speed up reads within the multi-level structure. Furthermore, LSMGraph incorporates a vertex-grained version control mechanism to mitigate the impact of LSM-tree compaction on read performance and ensure the correctness of concurrent read and write operations. Our evaluation shows that LSMGraph significantly outperforms state-of-the-art (graph) storage systems on both graph update and graph analytical workloads.

Graph Neural Networks (GNNs) have shown exceptional performance across a wide range of applications. Current frameworks leverage CPU-GPU heterogeneous environments for GNN model training, incorporating mini-batch and sampling techniques to mitigate GPU memory constraints. In such settings, sample-based GNN training can be divided into three phases: sampling, gathering, and training. Existing GNN systems deploy various task orchestration methods to execute each phase on either the CPU or GPU. However, through comprehensive experimentation and analysis, we observe that these task orchestration approaches do not optimally exploit the available heterogeneous resources, hindered by either inefficient CPU processing or GPU resource bottlenecks. In this paper, we propose NeutronOrch, a system for sample-based GNN training that ensures balanced utilization of the CPU and GPU. NeutronOrch decouples the training process by layer and pushes down the training task of the bottom layer to the CPU. This significantly reduces the computational load and memory footprint of GPU training. To avoid inefficient CPU processing, NeutronOrch only offloads the training of frequently accessed vertices to the CPU and lets GPU reuse their embeddings with bounded staleness. Furthermore, NeutronOrch provides a fine-grained pipeline design for the layer-based task orchestrating method. The experimental results show that compared with the state-of-the-art GNN systems, NeutronOrch can achieve up to 11.51× performance speedup.

Graph neural networks (GNNs) have emerged as a promising direction. Training large-scale graphs that relies on distributed computing power poses new challenges. Existing distributed GNN systems leverage data parallelism by partitioning the input graph and distributing it to multiple workers. However, due to the irregular nature of the graph structure, existing distributed approaches suffer from unbalanced workloads and high overhead in managing cross-worker vertex dependencies. In this paper, we leverage tensor parallelism for distributed GNN training. GNN tensor parallelism eliminates cross-worker vertex dependencies by partitioning features instead of graph structures. Different workers are assigned training tasks on different feature slices with the same dimensional size, leading to a complete load balance. We achieve efficient GNN tensor parallelism through two critical functions. Firstly, we employ a generalized decoupled training framework to decouple NN operations from graph aggregation operations, significantly reducing the communication overhead caused by NN operations which must be computed using complete features. Secondly, we employ a memory-efficient task scheduling strategy to support the training of large graphs exceeding single GPU memory, while further improving performance by overlapping communication and computation. By integrating the above techniques, we propose a distributed GNN training system NeutronTP. Our experimental results on a 16-node Aliyun cluster demonstrate that NeutronTP achieves 1.29×-8.72× speedup over state-of-the-art GNN systems including DistDGL, NeutronStar, and Sancus.

An efficient data structure is fundamental to meeting the growing demands in dynamic graph processing. However, the dual requirements for graph computation efficiency (with contiguous structures) and graph update efficiency (with linked list-like structures) present a conflict in the design principles of graph structures. After experimental studies of state-of-the-art dynamic graph structures, we observe that the overhead of cache misses accounts for a major portion of the graph computation time. This paper presents GastCoCo, a system with graph storage and coroutine-based prefetch co-design. By employing software prefetching via stackless coroutines and designing a prefetch-friendly data structure CBList, GastCoCo significantly alleviates the performance degradation caused by cache misses. Our results show that GastCoCo outperforms state-of-the-art graph storage systems by 1.3× - 180× in graph updates and 1.4× - 41.1× in graph computation.

Dynamic Graph Neural Networks (DGNNs) have demonstrated exceptional performance at dynamic-graph analysis tasks. However, the costs exceed those incurred by other learning tasks, to the point where deployment on large-scale dynamic graphs is infeasible. Existing distributed frameworks that facilitate DGNN training are in their early stages and experience challenges such as communication bottlenecks, imbalanced workloads, and GPU memory overflow. We introduce DynaHB, a distributed framework for DGNN training using so-called Hybrid Batches. DynaHB reduces communication by means of vertex caching, and it ensures even data and workload distribution by means of load-aware vertex partitioning. DyanHB also features a novel hybrid-batch training mode that combines vertex-batch and snapshot-batch techniques, thereby reducing training time and GPU memory usage. Next, to further enhance the hybrid batch based approach, DynaHB integrates a reinforcement learning-based batch adjuster and a pipelined batch generator with a batch reservoir to reduce the cost of generating hybrid batches. Extensive experiments show that DynaHB is capable of up to a 93× and an average of 8.06× speedups over the state-of-the-art training framework.

Many Graph Neural Network (GNN) training systems have emerged recently to support efficient GNN training. Since GNNs embody complex data dependencies between training samples, the training of GNNs should address distinct challenges different from DNN training in data management, such as data partitioning, batch preparation for mini-batch training, and data transferring between CPUs and GPUs. These factors, which take up a large proportion of training time, make data management in GNN training more significant. This paper reviews GNN training from a data management perspective and provides a comprehensive analysis and evaluation of the representative approaches. We conduct extensive experiments on various benchmark datasets and show many interesting and valuable results. We also provide some practical tips learned from these experiments, which are helpful for designing GNN training systems in the future.

Nowadays, studies on energy-efficient datacenters, especially the DataCenters powered by Renewable Energy mix (DCRE), have gained great attention. DCREs are large-scale, geo-distributed, and equipped with on-site renewable energy generators. For these features, it is expensive to perform empirical evaluations of proposed algorithms and solutions on the real-world DCREs, while the state-of-the-art datacenter simulators are not applicable for DCREs. In this paper, we present CloudSimPer (CLOUD SIMulator hybrid-Powered by rEnewable eneRgy), a general-purpose simulator that comprehensively supports the simulation of DCREs. Besides the functions such as renewable energy, geo-distribution, and long-term simulation, we also design evaluation metrics and an integrated simulation case for experimental studies in the future. The main challenge of CloudSimPer lies in designing a new model and software layer upon CloudSim, to solve the complexity of traceable and comparable simulations which connect renewable energies, datacenters, workloads, regions, and schedulers. We use the term schedulers broadly, encompassing any optimization approaches on DCREs for energy saving. We prove CloudSimPer and integrated case to be valid, so that simulation results are scientifically sound, by examining the expectation and the simulation results, and comparing the simulation results with selected competitors. CloudSimPer offers simulation services to datacenter designers, datacenter administrators, and academics.

Iterative graph processing is widely used as a significant paradigm for large-scale data analysis. In many global businesses of multinational enterprises, graph-structure data is usually geographically distributed in different regions to support low-latency services. Geo-distributed graph processing suffers from the Wide Area Networks (WANs) with scarce and heterogeneous bandwidth, thus essentially differs from traditional distributed graph processing. In this paper, we propose RAGraph, aRegion-Aware framework for geo-distributedgraphprocessing. At the core of RAGraph, we design a region-aware graph processing framework that allows advancing inefficient global updates locally and enables sensible coordination-free message interactions and flexible replaceable communication module. In terms of graph data preprocessing, RAGraph introduces a contribution-driven edge migration algorithm to effectively utilize network resources. RAGraph also contains an adaptive hierarchical message interaction engine to switch interaction modes adaptively based on network heterogeneity and fluctuation, and a discrepancy-aware message filtering strategy to filter important messages. Experimental results show that RAGraph can achieve an average speedup of 9.7× (up to 98×) and an average WAN cost reduction of 78.5$\%$(up to 97.3$\%$) compared with state-of-the-art systems.

Processing large graphs with memory-limited GPU needs to resolve issues of host-GPU data transfer, which is a key performance bottleneck. Existing GPU-accelerated graph processing frameworks reduce the data transfers by managing the active subgraph transfer at runtime. Some frameworks adopt explicit transfer management approaches based on explicit memory copy with filter or compaction. In contrast, others adopt implicit transfer management approaches based on on-demand access with zero-copy or unified-memory. Having made intensive analysis, we find that as the active vertices evolve, the performance of the two approaches varies in different workloads. Due to heavy redundant data transfers, high CPU compaction overhead, or low bandwidth utilization, adopting a single approach often results in suboptimal performance.In this work, we propose a hybrid transfer management approach to take the merits of both the two approaches at runtime, with an objective to achieve the shortest execution time in each iteration. Based on the hybrid approach, we present HyTGraph, a GPU-accelerated graph processing framework, which is empowered by a set of effective task scheduling optimizations to improve the performance. Our experimental results on real-world and synthesized graphs demonstrate that HyTGraph achieves up to 10.27X speedup over existing GPU-accelerated graph processing systems including Grus, Subway, and EMOGI.

Real-world graphs are constantly evolving, which demands updates of the previous analysis results to accommodate graph changes. By using the memoized previous computation state, incremental graph computation can reduce unnecessary recomputation. However, a small change may propagate over the whole graph and lead to large-scale iterative computations. To address this problem, we propose Layph, a two-layered graph framework. The upper layer is a skeleton of the graph which is much smaller than the original graph, and the lower layer has some disjoint subgraphs. Layph limits costly global iterative computations on the original graph to the small graph skeleton and a few subgraphs updated with the input graph changes. In this way, many vertices and edges are not involved in iterative computations, which significantly reduces the computation overhead and improves the performance of incremental graph processing. Our experimental results show that Layph outperforms current state-of-the-art incremental graph systems by 9.08× on average (up to 36.66×) in response time.

Distributed computing is promising to enable large-scale graph neural network (GNN) model training. However, care is needed to avoid excessive computational and communication overheads. Sampling is promising in terms of enabling scalability, and sampling techniques have been proposed to reduce training costs. However, online sampling introduces large overheads, and while offline sampling that is done only once can eliminate such overheads, it instead introduces information loss and accuracy degradation. Thus, existing sampling techniques are unable to improve simultaneously both efficiency and accuracy, particularly at low sampling rates. We develop a distributed system, ADGNN, for full-batch based GNN training that adopts a hybrid sampling architecture to enable a trade-off between efficiency and accuracy. Specifically, ADGNN employs sampling result reuse techniques to reduce the cost associated with sampling and thus improve training efficiency. To alleviate accuracy degradation, we introduce a new metric,Aggregation Difference (AD), that quantifies the gap between sampled and full neighbor set aggregation. We present so-called AD-Sampling that aims to minimize the Aggregation Difference with an adaptive sampling frequency tuner. Finally, ADGNN employs anAD -importance-based sampling technique for remote neighbors to further reduce communication costs. Experiments on five real datasets show that ADGNN is able to outperform the state-of-the-art by up to nearly 9 times in terms of efficiency, while achieving comparable accuracy to the non-sampling methods.

Multinational enterprises conduct global business that has a demand for geo-distributed transactional databases. Existing state-of-the-art databases adopt a sharded master-follower replication architecture. However, the single-master serving mode incurs massive cross-region writes from clients, and the sharded architecture requires multiple round-trip acknowledgments (e.g., 2PC) to ensure atomicity for cross-shard transactions. These limitations drive us to seek yet another design choice. In this paper, we propose a strongly consistent OLTP database GeoGauss with full replica multi-master architecture. To efficiently merge the updates from different master nodes, we propose a multi-master OCC that unifies data replication and concurrent transaction processing. By leveraging an epoch-based delta state merge rule and the optimistic asynchronous execution, GeoGauss ensures strong consistency with light-coordinated protocol and allows more concurrency with weak isolation, which are sufficient to meet our needs. Our geo-distributed experimental results show that GeoGauss achieves 7.06X higher throughput and 17.41X lower latency than the state-of-the-art geo-distributed database CockroachDB on the TPC-C benchmark.

Existing Graph Neural Network (GNN) training frameworks have been designed to help developers easily create performant GNN implementations. However, most existing GNN frameworks assume that the input graphs are static, but ignore that most real-world graphs are constantly evolving. Though many dynamic GNN models have emerged to learn from evolving graphs, the training process of these dynamic GNNs is dramatically different from traditional GNNs in that it captures both the spatial and temporal dependencies of graph updates. This poses new challenges for designing dynamic GNN training frameworks. First, the traditional batched training method fails to capture real-time structural evolution information. Second, the time-dependent nature makes parallel training hard to design. Third, it lacks system supports for users to efficiently implement dynamic GNNs. In this paper, we present NeutronStream, a framework for training dynamic GNN models. NeutronStream abstracts the input dynamic graph into a chronologically updated stream of events and processes the stream with an optimized sliding window to incrementally capture the spatial-temporal dependencies of events. Furthermore, NeutronStream provides a parallel execution engine to tackle the sequential event processing challenge to achieve high performance. NeutronStream also integrates a built-in graph storage structure that supports dynamic updates and provides a set of easy-to-use APIs that allow users to express their dynamic GNNs. Our experimental results demonstrate that, compared to state-of-the-art dynamic GNN implementations, NeutronStream achieves speedups ranging from 1.48X to 5.87X and an average accuracy improvement of 3.97%.

In many global businesses of multinational enterprises, graph-structure data is usually geographically distributed in different regions to support low-latency services. Geo-distributed graph processing suffers from the Wide Area Networks (WANs) with scarce and heterogeneous bandwidth, thus essentially differs from traditional distributed graph processing. In this paper, we propose RAGraph, a Region-Aware framework for geo-distributed graph processing. At the core of RAGraph, we design a region-aware graph processing framework that allows advancing inefficient global updates locally and enables sensible coordination-free message interactions. RAGraph also contains an adaptive hierarchical message interaction engine to switch interaction modes adaptively based on network heterogeneity and fluctuation, and a discrepancy-aware message filtering strategy to filter important messages. Finally, the experiments show that RAGraph can achieve 1.69X - 40.53X speedup and 20.9% - 97% WAN cost reduction compared with state-of-the-art systems.

GNN's training needs to resolve issues of vertex dependencies, i.e., each vertex representation's update depends on its neighbors. Existing distributed GNN systems adopt either a dependencies-cached approach or a dependencies-communicated approach. Having made intensive experiments and analysis, we find that a decision to choose one or the other approach for the best performance is determined by a set of factors, including graph inputs, model configurations, and an underlying computing cluster environment. If various GNN trainings are supported solely by one approach, the performance results are often suboptimal. We study related factors for each GNN training before its execution to choose the best-fit approach accordingly. We propose a hybrid dependency-handling approach that adaptively takes the merits of the two approaches at runtime. Based on the hybrid approach, we further develop a distributed GNN training system called NeutronStar, which makes high performance GNN trainings in an automatic way. NeutronStar is also empowered by effective optimizations in CPU-GPU computation and data processing. Our experimental results on 16-node Aliyun cluster demonstrate that NeutronStar achieves 1.81X-14.25X speedup over existing GNN systems including DistDGL and ROC.

Blockchain serves as a replicated transactional processing system in a trustless distributed environment. Existing blockchain systems all rely on an explicit ordering step to determine the global order of transactions that are collected from multiple peers. The ordering consensus can be the bottleneck since it must be Byzantine-fault tolerant and can scarcely benefit from parallel execution. In this paper, we propose an ordering-free architecture that makes ordering implicit through deterministic execution. Based on this novel architecture, we develop a permissioned blockchain system NeuChain. A number of key optimizations such as asynchronous block generation and pipelining are leveraged for high throughput and low latency. Several security mechanisms are also designed to make our system robust to malicious attacks. Our geo-distributed experimental results show that NeuChain can achieve 47.2--64.1X throughput improvement over HyperLedger Fabric and 1.6--12.2X throughput improvement over the state-of-the-art high performance blockchains.

The ever-growing amount of dynamic graph data demands efficient techniques of incremental graph processing. However, incremental graph algorithms are challenging to develop. Existing approaches usually require users to manually design nontrivial incremental operators, or choose different memoization strategies for certain specific types of computation, limiting the usability and generality. In light of these challenges, we propose Ingress, an automated system for incremental graph processing. Ingress is able to incrementalize batch vertex-centric algorithms into their incremental counterparts as a whole, without the need of redesigned logic or data structures from users. Underlying Ingress is an automated incrementalization framework equipped with four different memoization policies, to support all kinds of vertex-centric computations with optimized memory utilization. We identify sufficient conditions for the applicability of these policies. Ingress chooses the best-fit policy for a given algorithm automatically by verifying these conditions. In addition to the ease-of-use and generalization, Ingress outperforms state-of-the-art incremental graph systems by 15.93X on average (up to 147.14X) in efficiency.

Prioritized computation is shown promising performance for a large class of graph algorithms. It prioritizes the execution of some vertices that play important roles in determining convergence. For large-scale distributed graph processing, graph partitioning is an important preprocessing step that aims to balance workload and to reduce communication costs between workers. However, existing graph partitioning methods are designed for round-robin synchronous distributed frameworks. They balance workload without distinction of vertex importance and fail to consider the characteristics of priority-based scheduling, which may limit the benefit of prioritized graph computation. In this article, to accelerate prioritized iterative graph computations, we propose Hotness Balanced Partition (HBP). In prioritized graph computation, high priority vertices are likely to be executed more frequently and are likely to pass more messages, which result in hot vertices. Based on this observation, we partition graph by distributing vertices with distinction according to their hotness rather than blindly distributing vertices with equal weights, aiming to evenly distribute the hot vertices among workers. We further provide two HBP algorithms: a streaming-based algorithm for efficient one-pass processing and a distributed algorithm for distributed processing. Our results show that our proposed partitioning methods outperform the state-of-the-art partitioning methods, Fennel, HotGraph, and SNE.

Existing graph partition methods are designed for round-robin synchronous distributed frameworks. They balance workload without discrimination of vertex importance and fail to consider the characteristics of priority-based scheduling, which may limit the benefit of prioritized graph computation. To accelerate prioritized iterative graph computations, we propose Hotness Balanced Partition (HBP) and a stream-based partition algorithm Pb-HBP. Pb-HBP partitions graph by distributing vertices with discrimination according to their hotness rather than blindly distributing vertices with equal weights, which aims to evenly distribute the hot vertices among workers. Our results show that our proposed partition method outperforms the state-of-the-art partition methods, Fennel and HotGraph. Specifically, Pb-HBP can reduce 40-90% runtime of that by hash partition, 5-75% runtime of that by Fennel, and 22-50% runtime of that by HotGraph.

In database and large-scale data analytics, recursive aggregate processing plays an important role, which is generally implemented under a framework of incremental computing and executed synchronously and/or asynchronously. We identify three barriers in existing recursive aggregate data processing. First, the processing scope is largely limited to monotonic programs. Second, checking on conditions for monotonicity and correctness for async processing is sophisticated and manually done. Third, execution engines may be suboptimal due to separation of sync and async execution. In this paper, we lay an analytical foundation for conditions to check if a recursive aggregate program that is monotonic or even non-monotonic can be executed incrementally and asynchronously with its correct result. We design and implement a condition verification tool that can automatically check if a given program satisfies the conditions. We further propose a unified sync-async engine to execute these programs for high performance. To integrate all these effective methods together, we have developed a distributed Datalog system, called PowerLog. Our evaluation shows that PowerLog can outperform three representative Datalog systems on both monotonic and non-monotonic recursive programs.

In general, the performance of parallel graph processing is determined by three pairs of critical parameters, namely synchronous or asynchronous execution mode (Sync or Async), Push or Pull communication mechanism (Push or Pull), and Data-driven or Topology-driven traversing scheme (DD or TD), which increases the complexity and sophistication of programming and system implementation of GPU. Existing graph-processing frameworks mainly use a single combination in the entire execution for a given application, but we have observed their variable and suboptimal performance. In this paper, we present SEP-Graph, a highly efficient software framework for graph-processing on GPU. The hybrid execution mode is automatically switched among three pairs of parameters, with an objective to achieve the shortest execution time in each iteration. We also apply a set of optimizations to SEP-Graph, considering the characteristics of graph algorithms and underlying GPU architectures. We show the effectiveness of SEP-Graph based on our intensive and comparative performance evaluation on NVIDIA 1080, P100, and V100 GPUs. Compared with existing and representative GPU graph-processing framework Groute and Gunrock, SEP-Graph can reduce execution time up to 45.8 times and 39.4 times.

Increasingly more iterative and recursive query tasks are processed in data management systems, such as graph-structured data analytics, demanding fast response time. However, existing CTE-based recursive SQL and its implementation ineffectively respond to this intensive query processing with two major drawbacks. First, its iteration execution model is based on implicit set-oriented terminating conditions that cannot express aggregation-based tasks, such as PageRank. Second, its synchronous execution model cannot perform asynchronous computing to further accelerate execution in parallel. To address these two issues, we have designed and implemented SQLoop, a framework that extends the semantics of current SQL standard in order to accommodate iterative SQL queries. SQLoop interfaces between users and different database engines with two powerful components. First, it provides an uniform SQL expression for users to access any database engine so that they do not need to write database dependent SQL or move datasets from a target engine to process in their own sites. Second, SQLoop automatically parallelizes iterative queries that contain certain aggregate functions in both synchronous and asynchronous ways. More specifically, SQLoop is able to take advantage of intermediate results generated between different iterations and to prioritize the execution of partitions that accelerate the query processing. We have tested and evaluated SQLoop by using three popular database engines with real-world datasets and queries, and shown its effectiveness and high performance.

LSM-tree has been widely used in data management production systems for write-intensive workloads. However, as read and write workloads co-exist under LSM-tree, data accesses can experience long latency and low throughput due to the interferences to buffer caching from the compaction, a major and frequent operation in LSM-tree. After a compaction, the existing data blocks are reorganized and written to other locations on disks. As a result, the related data blocks that have been loaded in the buffer cache are invalidated since their referencing addresses are changed, causing serious performance degradations. To re-enable high-speed buffer caching during intensive writes, we propose Log-Structured buffered-Merge tree (simplified as LSbM-tree) by adding a compaction buffer on disks to minimize the cache invalidations on buffer cache caused by compactions. The compaction buffer efficiently and adaptively maintains the frequently visited datasets. In LSbM, strong locality objects can be effectively kept in the buffer cache with minimum or no harmful invalidations. With the help of a small on-disk compaction buffer, LSbM achieves a high query performance by enabling effective buffer caching, while retaining all the merits of LSM-tree for write-intensive data processing and providing high bandwidth of disks for range queries. We have implemented LSbM based on LevelDB. We show that with a standard buffer cache and a hard disk, LSbM can achieve 2x performance improvement over LevelDB. We have also compared LSbM with other existing solutions to show its strong cache effectiveness.

LSM-tree has been widely used in data management production systems for write-intensive workloads. However, as read and write workloads co-exist under LSM-tree, data accesses can experience long latency and low throughput due to the interferences to buffer caching from the compaction, a major and frequent operation in LSM-tree. After a compaction, the existing data blocks are reorganized and written to other locations on disks. As a result, the related data blocks that have been loaded in the buffer cache are invalidated since their referencing addresses are changed, causing serious performance degradations. In order to re-enable high-speed buffer caching during intensive writes, we propose Log-Structured buffered-Merge tree (simplified as LSbM-tree) by adding a compaction buffer on disks, to minimize the cache invalidations on buffer cache caused by compactions. The compaction buffer efficiently and adaptively maintains the frequently visited data sets. In LSbM, strong locality objects can be effectively kept in the buffer cache with minimum or without harmful invalidations. With the help of a small on-disk compaction buffer, LSbM achieves a high query performance by enabling effective buffer caching, while retaining all the merits of LSM-tree for write-intensive data processing, and providing high bandwidth of disks for range queries. We have implemented LSbM based on LevelDB. We show that with a standard buffer cache and a hard disk, LSbM can achieve 2x performance improvement over LevelDB. We have also compared LSbM with other existing solutions to show its strong effectiveness.

Density Peaks (DP) is a recently proposed clustering algorithm that has distinctive advantages over existing clustering algorithms. It has already been used in a wide range of applications. However, DP requires computing the distance between every pair of input points, therefore incurring quadratic computation overhead, which is prohibitive for large data sets. In this paper, we propose an efficient distributed algorithm LSHDDP, which is an approximate algorithm that exploits Locality Sensitive Hashing. We present formal analysis of LSH-DDP, and show that the approximation quality and the runtime can be controlled by tuning the parameters of LSH-DDP. Experimental results on both a local cluster and EC2 show that LSH-DDP achieves a factor of 1.7-70x speedup over the näıve distributed DP implementation and 2x speedup over the state-of-the-art EDDPC approach, while returning comparable cluster results.

Stream clustering is a fundamental problem in many streaming data analysis applications. Comparing to classical batch-mode clustering, there are two key challenges in stream clustering: (i) Given that input data are changing continuously, how to incrementally update their clustering results efficiently? (ii) Given that clusters continuously evolve with the evolution of data, how to capture the cluster evolution activities? Unfortunately, most of existing stream clustering algorithms can neither update the cluster result in real-time nor track the evolution of clusters. In this paper, we propose a stream clustering algorithm EDMStream by exploring the Evolution of Density Mountain. The density mountain is used to abstract the data distribution, the changes of which indicate data distribution evolution. We track the evolution of clusters by monitoring the changes of density mountains. We further provide efficient data structures and filtering schemes to ensure that the update of density mountains is in real-time, which makes online clustering possible. The experimental results on synthetic and real datasets show that, comparing to the state-of-the-art stream clustering algorithms, e.g., D-Stream, DenStream, DBSTREAM and MR-Stream, our algorithm is able to response to a cluster update much faster (say 7-15x faster than the best of the competitors) and at the same time achieve comparable cluster quality. Furthermore, EDMStream successfully captures the cluster evolution activities.

As new data and updates are constantly arriving, the results of data mining applications become stale and obsolete over time. Incremental processing is a promising approach to refresh mining results. It utilizes previously saved states to avoid the expense of re-computation from scratch. In this paper, we propose i2MapReduce, a novel incremental processing extension to MapReduce. Compared with the state-of-the-art work on Incoop, i2MapReduce (i) performs key-value pair level incremental processing rather than task level re-computation, (ii) supports not only one-step computation but also more sophisticated iterative computation, and (iii) incorporates a set of novel techniques to reduce I/O overhead for accessing preserved fine-grain computation states. Experimental results on Amazon EC2 show significant performance improvements of i2MapReduce compared to both plain and iterative MapReduce performing re-computation.

Density Peaks (DP) is a recently proposed clustering algorithm that has distinctive advantages over existing clustering algorithms. It has already been used in a wide range of applications. However, DP requires computing the distance between every pair of input points, therefore incurring quadratic computation overhead, which is prohibitive for large data sets. In this paper, we study efficient distributed algorithms for DP. We first show that a naive MapReduce solution (Basic-DDP) has high communication and computation overhead. Then, we propose LSH-DDP, an approximate algorithm that exploits Locality Sensitive Hashing for partitioning data, performs local computation, and aggregates local results to approximate the final results. We address several challenges in employing LSH for DP. We leverage the characteristics of DP to deal with the fact that some of the result values cannot be directly approximated in local partitions. We present formal analysis of LSH-DDP, and show that the approximation quality and the runtime can be controlled by tuning the parameters of LSH-DDP. Experimental results on both a local cluster and EC2 show that LSH-DDP achieves a factor of 1.7-70x speedup over the naive Basic-DDP and 2x speedup over the state-of-the-art EDDPC approach, while returning comparable cluster results. Compared to the popular K-means clustering, LSH-DDP also has comparable or better performance. Furthermore, LSH-DDP could achieve even higher efficiency with a lower accuracy requirement.

As new data and updates are constantly arriving, the results of data mining applications become stale and obsolete over time. Incremental processing is a promising approach to refreshing mining results. It utilizes previously saved states to avoid the expense of re-computation from scratch. In this paper, we propose i2MapReduce, a novel incremental processing extension to MapReduce, the most widely used frameworkfor mining big data. Compared with the state-of-the-art work on Incoop, i2MapReduce (i) performs key-value pair level incremental processing rather than task level re-computation, (ii) supports not only one-step computation but also more sophisticated iterative computation, which is widely used in data mining applications, and (iii) incorporates a set of novel techniques to reduce I/O overhead for accessing preserved fine-grain computation states. We evaluate i2MapReduce using a one-step algorithm and four iterative algorithms with diverse computation characteristics. Experimental results on Amazon EC2 show significant performance improvements of i2MapReduce compared to both plain and iterative MapReduce performing re-computation.

Myriad of graph-based algorithms in machine learning and data mining require parsing relational data iteratively. These algorithms are implemented in a large-scale distributed environment to scale to massive data sets. To accelerate these large-scale graph-based iterative computations, we propose delta-based accumulative iterative computation (DAIC). Different from traditional iterative computations, which iteratively update the result based on the result from the previous iteration, DAIC updates the result by accumulating the “changes” between iterations. By DAIC, we can process only the “changes” to avoid the negligible updates. Furthermore, we can perform DAIC asynchronously to bypass the high-cost synchronous barriers in heterogeneous distributed environments. Based on the DAIC model, we design and implement an asynchronous graph processing framework, Maiter. We evaluate Maiter on local cluster as well as on Amazon EC2 Cloud. The results show that Maiter achieves as much as 60 × speedup over Hadoop and outperforms other state-of-the-art frameworks.

Iterative computations are pervasive among data analysis applications, including web search, online social network analysis, recommendation systems, and so on. These applications typically involve data sets of massive scale. Fast convergence of the iterative computations on the massive data set is essential for these applications. In this paper, we explore the opportunity for accelerating iterative computations by prioritization. Instead of performing computations on all data points without discrimination, we prioritize the computations that help convergence the most, so that the convergence speed of iterative process is significantly improved. We develop a distributed computing framework, PrIter, which supports the prioritized execution of iterative computations. PrIter either stores intermediate data in memory for fast convergence or stores intermediate data in files for scaling to larger data sets. We evaluate PrIter on a local cluster of machines as well as on Amazon EC2 Cloud. The results show that PrIter achieves up to 50 × speedup over Hadoop for a series of iterative algorithms. In addition, PrIter is shown better performance for iterative computations than other state-of-the-art distributed frameworks such as Spark and Piccolo.

Expectation Maximization is a popular approach for parameter estimation in many applications such as image understanding, document classification, or genome data analysis. Despite the popularity of EM algorithms, it is challenging to efficiently implement these algorithms in a distributed environment. In particular, many EM algorithms that frequently update the parameters have been shown to be much more efficient than their concurrent counterparts. Accordingly, we propose two approaches to parallelize such EM algorithms in a distributed environment so as to scale to massive data sets. We prove that both approaches maintain the convergence properties of the EM algorithms. Based on the approaches, we design and implement a distributed framework, FreEM, to support the implementation of frequent updates for the EM algorithms. We show its efficiency through three well-known EM applications: k-means clustering, fuzzy c-means clustering and parameter estimation for the Gaussian Mixture model. We evaluate our framework on both a local cluster of machines and the Amazon EC2 cloud. Our evaluation shows that the EM algorithms with frequent updates implemented on FreEM can run much faster than those implementations with traditional concurrent updates.

Iterative computations are pervasive among data analysis applications in the cloud, including Web search, online social network analysis, recommendation systems, and so on. These cloud applications typically involve data sets of massive scale. Fast convergence of the iterative computation on the massive data set is essential for these applications. In this paper, we explore the opportunity for accelerating iterative computations and propose a distributed computing framework, PrIter, which enables fast iterative computation by providing the support of prioritized iteration. Instead of performing computations on all data records without discrimination, PrIter prioritizes the computations that help convergence the most, so that the convergence speed of iterative process is significantly improved. We evaluate PrIter on a local cluster of machines as well as on Amazon EC2 Cloud. The results show that PrIter achieves up to 50x speedup over Hadoop for a series of iterative algorithms.

Network virtualization is becoming an irresistible trend for building future Internet. It enables multiple virtual networks running on the same physical substrate, by which network service can be deployed on a customized virtual network to get better performance. In this paper, we propose a method to achieve more reliable live streaming service by deploying it over multiple virtual networks, namely MultiNet. Rather than running streaming service relying on a single network alone, MultiNet provides network resources redundancy for better reliability, avoiding unexpected interrupt due to network errors, such as link down, router crash, or transient routing failure. In addition, it balances traffic load between networks as some network congests, reducing packet losses significantly. To verify this idea, we implement MultiNet prototype exploiting some open source components. And several experiments on this prototype are studied both for illustrating MultiNet's benefits and for better understanding of MultiNet.

We focus on the problem of packet distribution for live media streaming over multipath networks. After proposing multipath live streaming model and an in-depth analysis of it, we suggest that traffic load should be allocated to paths in proportion to the paths' available bandwidths, considering complicated network status changing as well as burst of media sending rate, that can minimize the bandwidth overload probability. Moreover, weighted size-aware packet distribution algorithm is proposed to avoid the actual traffic deviation from expected due to variance of packet sizes. Simulation results show that the proposed algorithm outperforms other traditional algorithms, especially on reducing late packets, which has negative impact on real-time requirement.