高性能处理器中干扰公平队列I/O调度器  

An Interference Fair Queueing I/O Scheduler under High Performance Processors

作　　者：隋岩 叶嘉成 杨春[1] 佟冬[1] SUI Yan;YE Jiacheng;YANG Chun;TONG Dong(School of Electronics Engineering and Computer Science,Peking University,Beijing 100871)

机构地区：[1]北京大学信息科学技术学院,北京100871

出　　处：《北京大学学报（自然科学版）》2020年第6期1005-1012,共8页Acta Scientiarum Naturalium Universitatis Pekinensis

基　　金：国家科技重大专项(2018ZX01029101)资助。

摘　　要：高性能处理器和系统需要高存储带宽和高效的外部I/O处理,需要同时服务吞吐率密集应用和延迟敏感应用,给多道程序计算机系统和多租户模式的超级计算机系统的公平性带来巨大的挑战。针对这两类应用共享SSD(solid-state disks)等可并发的存储设备问题,开发一款基于队列的干扰公平(interference fair queueing,IFQ)调度器。在Linux操作系统实现IFQ调度器,并与其他调度器进行对比,包括Linux的CFQ调度器、STF调度器、MFAP的时间片流转调度器和MFAP的短时间片流转调度器。基于合成工作集、访问踪迹工作集和真实应用工作集的结果显示,IFQ调度器可以同时保证公平性和响应延迟。High performance processors and systems require high storage bandwidth and efficient external I/O processing.Many systems using high performance processors server a mixture of fully backlogged users,which continuously demand resources,and non-fully backlogged users.This presents challenges for fair resource management in multi-programmed computer systems and multi-tenant super computer systems.This paper develops a new storage I/O scheduler IFQ(interference fair queueing)for parallel accessible devices,such as SSDs.IFQ is implemented in Linux and compared with several existing I/O schedulers—Linux CFQ,STF,MFAP scheduler and MFAP scheduler with short time-slices.Results on synthetic I/O benchmarks,trace benchmarks,and real-world benchmarks demonstrate that only IFQ can achieve both fairness and high responsiveness.

关 键 词：存储调度器 干扰公平 固态硬盘 

分 类 号：TP332[自动化与计算机技术—计算机系统结构]