Journal of Parallel and Distributed Computing最新文献

Intrusion detection methods are crucial means to mitigate network security issues. However, the challenges posed by large-scale complex network environments include local information islands, regional privacy leaks, communication burdens, difficulties in handling heterogeneous data, and storage resource bottlenecks. Federated learning has the potential to address these challenges by leveraging widely distributed and heterogeneous data, achieving load balancing of storage and computing resources across multiple nodes, and reducing the risks of privacy leaks and bandwidth resource demands. This paper reviews the process of constructing federated learning based intrusion detection system from the perspective of intrusion detection. Specifically, it outlines six main aspects: application scenario analysis, federated learning methods, privacy and security protection, selection of classification models, data sources and client data distribution, and evaluation metrics, establishing them as key research content. Subsequently, six research topics are extracted based on these aspects. These topics include expanding application scenarios, enhancing aggregation algorithm, enhancing security, enhancing classification models, personalizing model and utilizing unlabeled data. Furthermore, the paper delves into research content related to each of these topics through in-depth investigation and analysis. Finally, the paper discusses the current challenges faced by research, and suggests promising directions for future exploration.

The incredible growth of Peer-to-Peer (P2P) networks has brought with it some complex challenges, such as trust issues and high bandwidth consumption. To address these challenges, this paper analyzes the “free-riding” behavior, system energy consumption, and the benefits of requesting and service peers in the network. A Geo/Geo/1+1 queuing model is built with malicious peers which includes several strategies such as repairable breakdown, synchronized multiple working vacations, differentiated service, and waiting threshold. The matrix-geometric solution method is used to obtain steady-state distribution and performance measures. By conducting numerical experiments and analyzing the impact of each parameter, it is possible to optimize the system's performance and reduce energy consumption. With careful adjustments to parameter values, significant cost savings of requesting peers and energy conservation can be achieved. The resulting analysis provides a comprehensive understanding of the behavior of P2P networks, and the strategies proposed in the study can be used to optimize the performance of P2P networks.

We propose a novel decentralized federated learning framework called B2DFL. It decomposes the aggregation process of vanilla FL into layered and serialized sub-aggregation processes and offloads the communication and computation from a single point to distributed nodes, thus addressing the single point of failure issue in centralized FL. The decentralization of B2DFL is based on the Butterfly, a distributed network topology, to organize and orchestrate the order and rules of node aggregation. Additionally, to mitigate potential risks such as dropouts or tampering, we leverage the blockchain and IPFS systems. Specifically, after each node completes its computation (including training and aggregation), it generates a hash value of the results as proof. We maintain a Tamper-evident Data Structure (TDS) on the blockchain, which records these proofs to ensure tamper-proofing and fast verification. To reduce the storage burden on the blockchain and improve throughput, we store the aggregated results on IPFS, a system that enables quick data location through hash values of data, for data backup. We also design a node replacement mechanism for quick dropout handling. We conduct a comprehensive performance evaluation and experimental results demonstrate that B2DFL presents a significant performance improvement while achieving privacy and decentralization.

Fortran's prominence in scientific computing requires strategies to ensure both that legacy codes are efficient on high-performance computing systems, and that the language remains attractive for the development of new high-performance codes. Coarray Fortran (CAF), part of the Fortran 2008 standard introduced for parallel programming, facilitates distributed memory parallelism with a syntax familiar to Fortran programmers, simplifying the transition from single-processor to multi-processor coding. This research focuses on innovating and refining a parallel programming methodology that fuses the strengths of Intel Coarray Fortran, Nvidia CUDA Fortran, and OpenMP for distributed memory parallelism, high-speed GPU acceleration and shared memory parallelism respectively. We consider the management of pageable and pinned memory, CPU-GPU affinity in NUMA multiprocessors, and robust compiler interfacing with speed optimisation. We demonstrate our method through its application to a parallelised Poisson solver and compare the methodology, implementation, and scaling performance to that of the Message Passing Interface (MPI), finding CAF offers similar speeds with easier implementation. For new codes, this approach offers a faster route to optimised parallel computing. For legacy codes, it eases the transition to parallel computing, allowing their transformation into scalable, high-performance computing applications without the need for extensive re-design or additional syntax.

Distributed computing, such as cloud computing, provides promising platforms for orchestrating scientific workflows' tasks based on their sequences and dependencies. Workflow scheduling plays an important role in optimizing concerned objectives for distributed computing, such as minimizing the makespan and cost. Many researchers have focused on optimizing a specific single workflow with multiple objectives. Currently, there are few studies on multi-workflow scheduling, with most research focusing on objectives such as cost and makespan. However, multi-workflow scheduling requires the design of specific objectives that reflect the unique characteristics of multiple workflows. On the other hand, clustering-based approaches have garnered significant attention in the field of workflow scheduling over distributed computing resources due to their advantage in reducing data communication among tasks. Despite this, the effectiveness of clustering-based algorithms has not been extensively studied and validated in the context of multi-objective multi-workflow scheduling models. Motivated by these factors, we propose an approach for multiple workflows' multi-objective optimization (MOO), considering the new defined metric, fairness. We first mathematically formulate the fairness and define a fairness-involved MOO model. Then, we propose an advanced clustering-based resource optimization strategy in multiple workflow runs. Experimental results show that the proposed approach performs better than the compared algorithms without significant compromise of the overall makespan and cost as well as individual fairness, which can guide the simulation workflow scheduling on clouds.

Graphs model several real-world phenomena. With the growth of unstructured and semi-structured data, parallelization of graph algorithms is inevitable. Unfortunately, due to inherent irregularity of computation, memory access, and communication, graph algorithms are traditionally challenging to parallelize. To tame this challenge, several libraries, frameworks, and domain-specific languages (DSLs) have been proposed to reduce the parallel programming burden of the users, who are often domain experts. However, existing frameworks to model graph algorithms typically target a single architecture. In this paper, we present a graph DSL, named StarPlat, that allows programmers to specify graph algorithms in a high-level format, but generates code for three different backends from the same algorithmic specification. In particular, the DSL compiler generates OpenMP for multi-core systems, MPI for distributed systems, and CUDA for many-core GPUs. Since these three are completely different parallel programming paradigms, binding them together under the same language is challenging. We share our experience with the language design. Central to our compiler is an intermediate representation which allows a common representation of the high-level program, from which individual backend code generations begin. We demonstrate the expressiveness of StarPlat by specifying four graph algorithms: betweenness centrality computation, page rank computation, single-source shortest paths, and triangle counting. Using a suite of ten large graphs, we illustrate the effectiveness of our approach by comparing the performance of the generated codes with that obtained with hand-crafted library codes. We find that the generated code is competitive to library-based codes in many cases. More importantly, we show the feasibility to generate efficient codes for different target architectures from the same algorithmic specification of graph algorithms.