面向SWF日志事件流数据的可视分析系统  

A Visual Analysis System for SWF Log Event Stream Data

作　　者：李玥 杨波 芦旭熠 单桂华[1] LI Yue;YANG Bo;LU Xuyi;SHAN Guihua(Computer Network Information Center,Chinese Academy of Sciences,Beijing 100190,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院计算机网络信息中心,北京100190 [2]中国科学院大学,北京100049

出　　处：《数据与计算发展前沿》2021年第4期3-17,共15页Frontiers of Data & Computing

基　　金：战略性先导科技专项(A类)(XDA19080102)。

摘　　要：【目的】为解决高性能计算集群的日志数据交互可视分析的问题,提供一套支持多种不同集群系统日志数据的可视分析框架,本文设计并实现了面向SWF日志数据的事件流可视分析系统SWFVis。【方法】SWFVis基于时间事件序列数据可视分析技术,结合多维可视分析、关联可视分析等技术,提出了一套针对SWF日志中的时序属性、关联属性、多维属性进行综合可视分析的方法。【结果】针对公开的iPSC/860与ForHLR II集群产生的真实数据,SWFVis直观展示了日志中整体作业的并发状态、执行情况,并对集群中存在的作业处理模式进行分析。【局限】对于事件流可视化呈现效果,在三维场景中存在的视角遮挡、分布杂乱的现象,未来考虑通过层次聚类等方法实现进一步的优化。【结论】SWFVis提供了针对SWF日志数据的交互式可视分析方法,支持多种计算机集群使用。通过对作业间存在的时序性、关联性进行展示,以及对作业多维属性进行挖掘,可以直观呈现集群工作状态、用户提交行为,并交互式发掘集群作业中存在的处理模式,为集群调度优化提供支撑。[Objective]HPC cluster log data is typically presented by traditional statistical charts,which has the disadvantage of no intuitive presentation and no interactivity.We design and implement SWFVis,an event stream visual analysis system for SWF log data to improve the presentation of the log data.[Methods]Based on the visual analysis technology of temporal event sequence data,combined with multi-dimensional visual analysis and associated visual analysis technology,SWFVis proposes a comprehensive visual analysis method for the sequential and multidimensional attributes of the jobs in the log event stream.[Results]Based on the real data generated by the public IPSC/860 and FORHLR II clusters,SWFVis intuitively displays the concurrent status and execution of the overall job in the logs,and analyzes the job processing patterns in the clusters.[Limitations]As for the visual rendering effect of event stream,there are still some problems in the three-dimensional scenes such as view blocking and disordered distribution.In the future,we consider achieving hierarchical clustering and other methods for further optimization.[Conclusions]SWFVis provides an interactive visual analysis method for SWF log data.It intuitively presents the working state of the cluster and user submission behavior,and supports interactive exploration of the processing pattern in the cluster operation,which provides support for the optimization of cluster scheduling.

关 键 词：高性能计算集群 时间事件序列 事件流 可视分析 SWF日志 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] O211.61[自动化与计算机技术—控制科学与工程]