面向大电网在线实时仿真的通信敏感资源调度  

Communication-aware gang scheduling for online real-time power grid simulations

作　　者：袁雨馨 唐宏伟[1] 赵晓芳[1] 严剑峰[3] 周二专 YUAN Yuxin;TANG Hongwei;ZHAO Xiaofang;YAN Jianfeng;ZHOU Erzhuan(Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190;University of Chinese Academy of Sciences,Beijing 100049;China Electric Power Research Institute,Beijing 100192)

机构地区：[1]中国科学院计算技术研究所,北京100190 [2]中国科学院大学,北京100049 [3]中国电力科学研究院,北京100192

出　　处：《高技术通讯》2022年第8期845-858,共14页Chinese High Technology Letters

基　　金：国家重点研发计划(2018YFB0904503)资助项目。

摘　　要：面向大电网的实时仿真计算被建模为协同运行的多进程并行计算任务的集合,任务调度与资源分配的优劣是其能否实现实时性目标的关键因素之一。本文针对大规模电网机电-电磁混合仿真计算的应用场景,根据任务进行了特性分析,总结了资源利用率规律,提出了一种通信敏感的组调度框架(CGS)。该框架提出了一种集中式两阶段调度架构,以在不中断在线流运行过程中对任务进行主动采样和调度,达到精准预测需求的目标;提出了基于通信图的图划分策略与基于调度模型的匹配策略相结合的CGS调度算法,实现了进程组调度,降低了任务跨节点的通信开销。实验从任务平均划分数、平均割边成本、有负载节点数、集群资源利用率和资源碎片率指标方面对5种基线算法与应用了CGS的算法进行对比测试,结果表明CGS与基本策略相比至少降低了37%的进程间通信开销,减少了19%的资源碎片,平均提高了34%的集群资源利用率。Real-time simulations of power grid are modeled as a collection of multi-process parallel tasks running in cooperation. The result of task scheduling and resource allocation is one of the key factors deciding whether a task can achieve its real-time goal. In this paper, a communication-aware gang scheduling framework(CGS) is proposed for the application scenario of large-scale grid electromechanical-electromagnetic hybrid simulation. CGS is based on task profiling and the resource utilization summary. CGS builds a centralized two-stage scheduling architecture to proactively sample and schedule tasks without interrupting online streaming process to predict accurate requirements;it proposes a CGS scheduling algorithm that combines a communication-based graph partitioning strategy with a model-based matching strategy to achieve gang scheduling and reduce the communication overhead of tasks on multi-nodes. Experiments are conducted to test five baseline algorithms against the algorithm with CGS applied in terms of five metrics: the average number of task partition, the average cost of edge-cut, the number of workload bearing nodes, the cluster resource utilization and the resource fragmentation rate. The experimental results conclusively demonstrate that CGS reduces the inter-process communication overhead by at least 37% and reduces the resource fragmentation by 19%, and also improves the cluster resource utilization by 34% on average compared to the baseline algorithms.

关 键 词：进程组调度 在线实时电网仿真 通信敏感 多维资源分配 

分 类 号：TM73[电气工程—电力系统及自动化]