High-performance Distributed Computing Research Articles

ServiceNet: resource-efficient architecture for topology discovery in large-scale multi-tenant clouds

AbstractModern computing infrastructures are evolving due to virtualisation, especially with the advent of 5G and future technologies. While this transition offers numerous benefits, it also presents challenges. Consequently, understanding these complex systems, including networks, services, and their interconnections, is crucial. This paper introduces ServiceNet, a groundbreaking architecture that accurately performs the important task of providing understanding of a multi-tenant architecture by discovering the complete topology, crucial in the realm of high-performance distributed computing. Experimental results have been carried out in different scenarios in order to validate our approach, demonstrating the effectiveness of our approach in comprehensive multi-tenant topology discovery. The experiments, involving up to forty tenant, highlight the adaptability of ServiceNet as a valuable tool for real-time monitoring in topology discovery purposes, even in challenging scenarios.

Retracted: Innovative Research on Urban Community Governance Decision-Making Relying on Distributed High-Performance Computing Blockchain Key Algorithms.

[This retracts the article DOI: 10.1155/2022/4078014.].

Test Method of Secure and Controllable FPGA Chip Used in Flexible DC Transmission Control and Protection System

Flexible DC transmission control and protection system is a high-performance distributed computing system, which has the characteristics of high-speed data acquisition and fast processing, and uses a large number of high-performance FPGA (Field Programmable Logic Array) chips. FPGA has the characteristics of high integration and complex manufacturing process, and its security and controllability is faced with many challenges. The reliability of FPGA is an important standard to measure its performance, so reliability test fault screening is very important to ensure its normal operation. Although the automatic test equipment (ATE) can test the fault of FPGA, it is not convenient for field testing due to the price, test environment and other factors. In this paper, a design scheme of built-in self-test (BIST) for fault detection and fault diagnosis is proposed, which can detect faults of look-up table (LUT) and interconnection resources (wires and programmable switches). At the same time, considering the effect of rising temperature on fault probability, a controlled self-heating element (SHE) is constructed by using the internal resources of FPGA to provide the required temperature for BIST testing. Finally, an integrated layout scheme of BIST and SHE is proposed. An experiment of self-heating of FPGA chip is completed by TSINGHUA PGL22G, which verifies the correctness and feasibility of the theory.

Dynamic Replication Policy on HDFS Based on Machine Learning Clustering

Data growth in recent years has been swift, leading to the emergence of big data science. Distributed File Systems (DFS) are commonly used to handle big data, like Google File System (GFS), Hadoop Distributed File System (HDFS), and others. The DFS should provide the availability of data and reliability of the system in case of failure. The DFS replicates the files in different locations to provide availability and reliability. These replications consume storage space and other resources. The importance of these files differs depending on how frequently they are used in the system. So some of these files do not deserve to replicate many times because it is unimportant in the system. This paper introduces a Dynamic Replication Policy using Machine Learning Clustering (DRPMLC) on HDFS, which uses Machine Learning to cluster the files into different groups and apply other replication policies to each group to reduce the storage consumption, improve the read and write operations time and keep the availability and reliability of HDFS as a High-Performance Distributed Computing (HPDC).

Task Scheduling Optimization in Cloud Computing by Coronavirus Herd Immunity Optimizer Algorithm

Cloud computing is now dominant in high-performance distributed computing, offering resource polling and ondemand services over the web. So, the task scheduling problem in a cloud computing environment becomes a significant analysis space due to the dynamic demand for user services. The primary goal of scheduling tasks is to allocate tasks to processors to achieve the shortest possible makespan while respecting priority restrictions. In heterogeneous multiprocessor systems, task and schedule assignments significantly impact the system's operation. Therefore, the different processes within the heuristic-based scheduling task algorithm will lead to a different makespan on a heterogeneous computing system. Thus, a suitable algorithm for scheduling should set precedence efficiently for every subtask based on the resources required to reduce its makespan. This paper proposes a novel efficient scheduling task algorithm based on the coronavirus herd immunity optimizer algorithm to solve task scheduling problems in a cloud computing environment. The basic idea of this method is to use the advantages of meta-heuristic algorithms to get the optimal solution. We evaluate the performance of our algorithm by applying it to three cases. The collected findings suggest that the proposed strategy successfully achieved the best solution in terms of makespan, speedup, efficiency, and throughput compared to others. Furthermore, the results demonstrate that the suggested technique beats existing methods new genetic algorithm (NGA), proposed particle swarm optimization (PPSO), whale optimization algorithm (WOA), enhanced genetic algorithm for task scheduling (EGA-TS), gravitational search algorithm (GSA), genetic algorithm (GA), and hybrid heuristic and genetic (HHG) by 22.8%, 12.3%, 8.8%, 7.3%, 7.3%, 3.4%, and 3.4% respectively according to makespan.

Adaptively restarted block Krylov subspace methods with low-synchronization skeletons

With the recent realization of exascale performance by Oak Ridge National Laboratory’s Frontier supercomputer, reducing communication in kernels like QR factorization has become even more imperative. Low-synchronization Gram-Schmidt methods, first introduced in Świrydowicz et al. (Numer. Lin. Alg. Appl. 28(2):e2343, 2020), have been shown to improve the scalability of the Arnoldi method in high-performance distributed computing. Block versions of low-synchronization Gram-Schmidt show further potential for speeding up algorithms, as column-batching allows for maximizing cache usage with matrix-matrix operations. In this work, low-synchronization block Gram-Schmidt variants from Carson et al. (Linear Algebra Appl. 638:150–195, 2022) are transformed into block Arnoldi variants for use in block full orthogonalization methods (BFOM) and block generalized minimal residual methods (BGMRES). An adaptive restarting heuristic is developed to handle instabilities that arise with the increasing condition number of the Krylov basis. The performance, accuracy, and stability of these methods are assessed via a flexible benchmarking tool written in MATLAB. The modularity of the tool additionally permits generalized block inner products, like the global inner product.

GARLSched: Generative adversarial deep reinforcement learning task scheduling optimization for large-scale high performance computing systems

Efficient task scheduling has become increasingly complex as the number and type of tasks proliferate and the size of computing resource grows in large-scale distributed high-performance computing (HPC) systems. At present, deep reinforcement learning (DRL) methods have achieved certain success in scheduling problems. However, due to the exogeneity of the task and the sparsity of the reward, the learning of the DRL control policy requires a significant amount of training time and data and cannot guarantee effective convergence. Meanwhile, based on the understanding of HPC system characteristics, various scheduling policies with acceptable performance for different optimization goals have been developed by the experts. But these heuristic methods cannot adapt to environmental changes and optimize for specific workloads. Therefore, the generative adversarial reinforcement learning scheduling (GARLSched) algorithm is proposed to effectively guide the learning of DRL in large-scale dynamic scheduling issues based on the optimal policy in the expert pool. In addition, the task embedding-based discriminator network effectively improves and stabilizes the learning process. Experiments show that compared with heuristic and DRL scheduling algorithms, GARLSched can learn high-quality scheduling policies for various workloads and optimization objects. Furthermore, the learned models can perform stably even when applied to invisible workloads, making them more practical in HPC systems.

Innovative Research on Urban Community Governance Decision-Making Relying on Distributed High-Performance Computing Blockchain Key Algorithms.

Since the country began to go global, the country's economy has developed rapidly, cultural exchanges between countries have become more and more frequent, and foreign cultures have begun to gradually spread to the country. At the same time, through the absorption of foreign community governance experience, domestic research on community governance has also begun to be put on the agenda. This paper aims to study the innovative exploration of urban community governance mechanisms through key high-performance computing algorithms. To this end, this paper proposes a combination of the Bayesian algorithm and distributed high-performance computing to analyze and explore the governance and management methods of the community through its efficient and stable computing power, and derive the most suitable community governance mechanism. An experiment was also set up for comparative analysis. The experimental results show that the community governance mechanism derived from the key distributed high-performance computing algorithm improves the community governance capability by 19.4%, effectively improving the community governance and management issues.

Task Scheduling Optimization in Cloud Computing by Rao Algorithm

Cloud computing is currently dominated within the space of high-performance distributed computing and it provides resource polling and on-demand services through the web. So, task scheduling problem becomes a very important analysis space within the field of a cloud computing environment as a result of user's services demand modification dynamically. The main purpose of task scheduling is to assign tasks to available processors to produce minimum schedule length without violating precedence restrictions. In heterogeneous multiprocessor systems, task assignments and schedules have a significant impact on system operation. Within the heuristic-based task scheduling algorithm, the different processes will lead to a different task execution time (makespan) on a heterogeneous computing system. Thus, a good scheduling algorithm should be able to set precedence efficiently for every subtask depending on the resources required to reduce (makespan). In this paper, we propose a new efficient task scheduling algorithm in cloud computing systems based on RAO algorithm to solve an important task and schedule a heterogeneous multiple processing problem. The basic idea of this process is to exploit the advantages of heuristic-based algorithms to reduce space search and time to get the best solution. We evaluate our algorithm's performance by applying it to three examples with a different number of tasks and processors. The experimental results show that the proposed approach significantly succeeded in finding the optimal solutions than others in terms of the time of task implementation.

Adaptive Workflow Scheduling to Increase Fault Tolerance in Cloud Computing

Cloud computing in the field of high-performance distributed computing has emerged as a new development in which the demand for access to resources via the Internet is presented in distributed servers that dynamically scale Are acceptable. One of the important research issues that must be considered to achieve efficient performance is fault tolerance. Fault tolerance is a way to find faults and failures in a system. Predicting and reducing errors play an important role in increasing the performance and popularity of cloud computing. In this study, an adaptive workflow scheduling approach is presented to increase fault tolerance in cloud computing. The present approach calculates the probability of failure for each resource according to the execution time of tasks on the resources. In the present method, a deadline is set for each of the tasks. If the task is not completed within the specified time, the probability of failure in the source increases and subsequent tasks are not sent to the desired source. The simulation results of the proposed method show that the proposed idea can work well on workflows and improve service quality factors.

Call for papers: Special issue on distributed next generation 5G data networks

The eminent growth in connected devices technologies, like the Internet of Things (IoT), 5 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">th</sup> generation (5G) and beyond communication systems, lead to innovations for human beings. The uninterrupted data communication services play diverse roles in the mobile user's routine life. The various changing user group patterns are unpredictable due to the day-by-day up-gradation of user devices. Even though the existing telecommunication services have upgraded up to 5G with high performance distributed computing (HPDC) in the distributed data networks (DDN) for ultra-reliable and low latency communication (URLLC) services, yet the backend distributed database systems (DDS) are facing serious security and privacy issues due to lack of federated intelligence in the data networks. Recently, we have seen an increased focus and effort by users and policymakers toward enhancing security and privacy related to the collection and usage of the data in DDNs. When it comes to the intelligence using machine/deep learning (ML/DL) of the HPDC systems for security and privacy, enough dataset is required, which often includes personal user information to train ML/DL models. As data privacy and security represents a growing critical concern, given the above-mentioned new areas of legislation and policies, novel ML methodologies like federated learning (FL) have been developed in part to address these concerns.

A new scalable distributed k-means algorithm based on Cloud micro-services for High-performance computing

The paper aims to propose a distributed clustering method for High performance computing (HPC) models and, its application for medical image processing. The communication cost is one of the great challenges, which minimizes the scalability of parallel and distributed computing models. Indeed, it reduces significantly the performance of HPC systems where these models are assigned to be implemented. In this paper, we present a new distributed k-means method which integrates virtual parallel distributed computing model with a low communication cost mechanism. The k-means method is performed as a distributed service within a cooperative micro-services team which uses asynchronous communication mechanism based on AMQP protocol. We design and implement a parallel and distributed HPC application for MRI image segmentation assigned to be deployed on cloud. Experimental results show that the proposed method (DSCM) and its assigned model reach high degree of scalability. We expect this clustering approach to provide scalable HPC applications for big data clustering.

A Survey of End-System Optimizations for High-Speed Networks

The gap is widening between the processor clock speed of end-system architectures and network throughput capabilities. It is now physically possible to provide single-flow throughput of speeds up to 100 Gbps, and 400 Gbps will soon be possible. Most current research into high-speed data networking focuses on managing expanding network capabilities within datacenter Local Area Networks (LANs) or efficiently multiplexing millions of relatively small flows through a Wide Area Network (WAN). However, datacenter hyper-convergence places high-throughput networking workloads on general-purpose hardware, and distributed High-Performance Computing (HPC) applications require time-sensitive, high-throughput end-to-end flows (also referred to as “elephant flows”) to occur over WANs. For these applications, the bottleneck is often the end-system and not the intervening network. Since the problem of the end-system bottleneck was uncovered, many techniques have been developed which address this mismatch with varying degrees of effectiveness. In this survey, we describe the most promising techniques, beginning with network architectures and NIC design, continuing with operating and end-system architectures, and concluding with clean-slate protocol design.

There is no McLaughlin geometry

We determine that there is no partial geometry G with parameters (s,t,α)=(4,27,2). The existence of such a geometry has been a challenging open problem of interest to researchers for almost 40 years. The particular interest in G is due to the fact that it would have the exceptional McLaughlin graph as its point graph. Our proof makes extensive use of symmetry and high-performance distributed computing, and details of our techniques and checks are provided. One outcome of our work is to show that a pseudogeometric strongly regular graph achieving equality in the Krein bound need not be the point graph of any partial geometry.

Stochastic scheduling for variation-aware virtual machine placement in a cloud computing CPS

As the most promising computing paradigm, cloud computing opens up new horizons for the area of high-performance distributed computing. Cyber-physical Systems (CPSs) present novel digital systems, which integrate computation, communication and the control of physical resource. Applied CPSs architecture in cloud computing can provide real-time and scalable resource monitoring and offer time-critical applications. With unrivaled scalability and flexibility, the CPSs based cloud services brings significant convenience to customers in need of elastic computing power. The quality of CPSs based cloud services is, to an large extent, determined by the performance of Virtual Machine (VM) placement algorithm for the data center. VM placement also effect the communication between applications and physical resource distribution in cloud computing CPSs.The traditional VM placement algorithm is built upon the two-tier architecture. With the presence of multi-media applications, the application level controller cannot accurately quantify the varying amount computing resources required by VMs at runtime. Consequently, lacking accurate resource demand for each VM, controller at the data center level cannot generate the VM placement with satisfactory feasibility. This architecture no longer fits the modern data centers. In this paper, the two tier VM placement framework is proposed to resolve this technical challenge. Our LP-based variation-unaware VM placement algorithm generates the VM placement with minimized energy consumption. On the other hand, our feasibility driven stochastic VM placement (FDSP) algorithm works seamlessly with the LP-based algorithm to achieve desirable feasibility of the placement. Our experimental results show that the LP-based variation unaware VM placement algorithm improves the energy consumption by 15.3% on average from the baseline algorithm. For test cases with resource request variations, the FDSP algorithm saves 15.7% energy cost compared to the “worst case scenario” of the traditional VM placement paradigm. On the other hand, it improves the feasibility by 50.0% compared to the “best case scenario”.

Exploring the Design Tradeoffs for Extreme-Scale High-Performance Computing System Software

Owing to the extreme parallelism and the high component failure rates of tomorrow's exascale, high-performance computing (HPC) system software will need to be scalable, failure-resistant, and adaptive for sustained system operation and full system utilizations. Many of the existing HPC system software are still designed around a centralized server paradigm and hence are susceptible to scaling issues and single points of failure. In this article, we explore the design tradeoffs for scalable system software at extreme scales. We propose a general system software taxonomy by deconstructing common HPC system software into their basic components. The taxonomy helps us reason about system software as follows: (1) it gives us a systematic way to architect scalable system software by decomposing them into their basic components; (2) it allows us to categorize system software based on the features of these components, and finally (3) it suggests the configuration space to consider for design evaluation via simulations or real implementations. Further, we evaluate different design choices of a representative system software, i.e. key-value store, through simulations up to millions of nodes. Finally, we show evaluation results of two distributed system software, Slurm++ (a distributed HPC resource manager) and MATRIX (a distributed task execution framework), both developed based on insights from this work. We envision that the results in this article help to lay the foundations of developing next-generation HPC system software for extreme scales.

Molecular simulation workflows as parallel algorithms: the execution engine of Copernicus, a distributed high-performance computing platform.

Computational chemistry and other simulation fields are critically dependent on computing resources, but few problems scale efficiently to the hundreds of thousands of processors available in current supercomputers-particularly for molecular dynamics. This has turned into a bottleneck as new hardware generations primarily provide more processing units rather than making individual units much faster, which simulation applications are addressing by increasingly focusing on sampling with algorithms such as free-energy perturbation, Markov state modeling, metadynamics, or milestoning. All these rely on combining results from multiple simulations into a single observation. They are potentially powerful approaches that aim to predict experimental observables directly, but this comes at the expense of added complexity in selecting sampling strategies and keeping track of dozens to thousands of simulations and their dependencies. Here, we describe how the distributed execution framework Copernicus allows the expression of such algorithms in generic workflows: dataflow programs. Because dataflow algorithms explicitly state dependencies of each constituent part, algorithms only need to be described on conceptual level, after which the execution is maximally parallel. The fully automated execution facilitates the optimization of these algorithms with adaptive sampling, where undersampled regions are automatically detected and targeted without user intervention. We show how several such algorithms can be formulated for computational chemistry problems, and how they are executed efficiently with many loosely coupled simulations using either distributed or parallel resources with Copernicus.

Node-Oriented Workflow (NOW): A Command Template Workflow Management Tool for High Throughput Data Analysis Pipelines

Next generation sequencing (NGS) systems produce vast quantities of data that require substantial computational resources for typical analysis tasks. In addition, data that are generated by different NGS systems are not homogeneous. Moreover, there are an overwhelming number of tools available for performing typical tasks. Managing NGS workflows involves writing custom scripts that quickly grow in complexity, often resulting in unwieldy workflows that underutilize typical high performance compute resources, and increase the demands of the staff managing these workflows. We present Node-Oriented Workflow (NOW), a dynamic command template workflow engine for high performance distributed computing (HPC) systems. Our system provides a simple-to-use browserbased front end for designing and managing complex workflows. Workflows are configured using a simple browser interface, and are managed by the integrated job engine, which initializes nodes, monitors node status, and processes results of individual jobs across nodes in an HPC configuration. We reduce excessive messaging across nodes by placing the burden on nodes to start tasks in a workflow when dependencies are met, i.e., node oriented workflow. Our system was designed for NGS processing in the clinical research setting, emphasizing user simplicity, tool scalability, minimization of redundancy in workflows, while maximizing throughput in an HPC environment. Furthermore, NOW is not restricted to NGS pipeline management, but can used to manage any computational pipeline.

An Adaptive Redundant Reservation Strategy in Distributed High-performance Computing Environments

In distributed high-performance computing environments, resource reservation mechanism is an effective approach to provide desirable quality of service for large-scale applications. However, conventional reservation service might result in lower resource utilization and higher rejection rate if it is excessively applied. Furthermore, redundant reservation policy has been widely applied in many practical systems with aiming to improve the reliability of application execution at runtime. In this paper, we proposed an adaptive redundant reservation strategy, which uses overlapping technique to implement reservation admission and enable resource providers dynamically determine the redundant degree at runtime. By overlapping a new reservation with an existing one, a request whose reservation requirements can not be satisfied in traditional way might be accepted. Also, by dynamically determining the redundant degree, our strategy can obtain optimal tradeoff between performance and reliability for distributed high-performance computing systems. Experimental results show that the strategy can bring about remarkably higher resource utilization and lower rejection rate when using redundant reservation service at the price of a slightly increasing of reservation violations.

From Quantity To Quality: Massive Molecular Dynamics Simulation of Nanostructures under Plastic Ddeformation in Desktop and Service Grid Distributed Computing Infrastructure

The distributed computing infrastructure (DCI) on the basis of BOINC and EDGeS-bridge technologies for high-performance distributed computing is used for porting the sequential molecular dynamics (MD) application to its parallel version for DCI with Desktop Grids (DGs) and Service Grids (SGs). The actual metrics of the working DG-SG DCI were measured, and the normal distribution of host performances, and signs of log-normal distributions of other characteristics (CPUs, RAM, and HDD per host) were found. The practical feasibility and high efficiency of the MD simulations on the basis of DG-SG DCI were demonstrated during the experiment with the massive MD simulations for the large quantity of aluminum nanocrystals ($\sim10^2$-$10^3$). Statistical analysis (Kolmogorov-Smirnov test, moment analysis, and bootstrapping analysis) of the defect density distribution over the ensemble of nanocrystals had shown that change of plastic deformation mode is followed by the qualitative change of defect density distribution type over ensemble of nanocrystals. Some limitations (fluctuating performance, unpredictable availability of resources, etc.) of the typical DG-SG DCI were outlined, and some advantages (high efficiency, high speedup, and low cost) were demonstrated. Deploying on DG DCI allows to get new scientific $\it{quality}$ from the simulated $\it{quantity}$ of numerous configurations by harnessing sufficient computational power to undertake MD simulations in a wider range of physical parameters (configurations) in a much shorter timeframe.