Towards Application IO Behavior Collection And Quantification Analysis On Sugon High Performance Computer

Nowadays,with the progress of science and technology,high performance computing has been widely used in many fields,such as climate simulation,hydrodynamics,molecular dynamics,bioinformatics and so on.Most of these applications are scientific computing applications,which are developing towards higher resolution and larger parallel scale.What can match these trend for high performance computers is that the computing power is constantly improving,but the development of their storage system is relatively slow,"storage wall"problem is more and more prominent,the speed between computing and 10 is more and more mismatched,and it is more and more difficult to design and build a balanced architecture.In the actual production environment,for many scientific computing applications,10 performance rather than computing performance in many scenarios has become a bottleneck to further improve its overall performance.Therefore,on the premise of not changing and upgrading the hardware and software architecture of high-performance computer storage system,from the perspective of file system and application IO,how to comprehensively and correctly understand the application IO behavior characteristics,quickly and accurately locate the application 10 hotspot,and alleviate and improve the "storage wall" problem has become an important research topic.In order to explore and understand the HPC application IO behavior and analyze the IO performance quantitatively,this paper designs,implements and deploys a set of 10 resource monitoring and analysis system--IOStar,which integrates IO behavior information collection,storage and visual analysis functions on a high performance computing cluster system.IOStar is deployed on Sugon high-performance computing cluster system.As a system tool application in real production envirorment,it provides quantitative IO analysis services for HPC applications,locates and analyzes 10 hotspots.IOStar consists of three subsystems:IOStar collection subsystem,IO log data storage subsystem and job IO analysis and visualization subsystem.The IOStar collection subsystem is responsible for collecting IO log information and job process information customized by the client of the shared file system on the computing node,and then sending IO log information to the IO log storage subsystem through Ethernet,sending job process information to the job 10 analysis subsystem;the IO log data storage subsystem is responsible for collecting IO log information data and persistently storing it to the distributed storage log collection and storage service cluster;the job IO analysis subsystem is responsible for quantitative analysis of IO behavior,characteristics and performance problems of applications,which is displayed in the form of visual analysis results.providing a powerful reference for analysis of application IO hotspot and optimization of application IO performance.Using the IO log data collected and stored by IOStar,this paper proposes a method of IO behavior analysis and performance diagnosis for HPC applications.Through the quantitative analysis of IO log information data,from the perspective of application,file and process,taking IO bandwidth,IOPs(IO operations per second),read-write data volume,file IO access mode and other quantitative indicators as analysis parameters,the IO behavior analysis and IO performance diagnosis of HPC application are carried out,and the application IO hotspot is analyzed.Through the work of this paper,we can accumulate experience for deploying a more comprehensive system resource monitoring and analysis platform on a larger scale and high performance computer system in the future,and provide optimization suggestions for the IO module design of HPC application program in the future.