一个支持访存带宽敏感调度的跨执行优化方法  

A Cross-Run Optimization Approach for Supporting Memory Bandwidth Aware Job Scheduling

作　　者：徐地[1] 武成岗[1] 冯晓兵[1] 

机构地区：[1]中国科学院计算技术研究所计算机体系结构国家重点实验室,北京100190

出　　处：《计算机学报》2014年第7期1580-1592,共13页Chinese Journal of Computers

基　　金：国家自然科学基金(60736012);国家"九七三"重点基础研究发展规划项目基金(2005CB321602);工信部核高基重大专项(2009ZX01036-001-002)资助~~

摘　　要：片外访存带宽是共享存储多核系统的主要性能瓶颈.访存带宽敏感的任务调度可以有效缓解并发程序间的访存竞争,提高系统吞吐率.然而调度策略的实施需要关于程序执行的先验知识,给系统用户增加了额外负担;另一方面,并发程序间的带宽竞争使得运行时收集的程序带宽需求信息不精确,影响了调度效果.在该文中,作者提出了一个低开销、对用户透明的跨执行优化方法解决上述问题.它在运行时识别程序的阶段性(phase)行为,并估算每个phase的独占执行性能;上述信息被存储到数据库中,在程序未来的执行中指导调度,并且信息精度随着程序的多次执行持续增加.上述过程使得带宽敏感调度策略的进行不再需要任何用户信息制导,并且优化了调度效果.作者在基于Intel Xeon处理器的8核系统上实现并评估了该系统,测试结果表明:相对于Linux操作系统(OS)默认的调度策略,该文的方法能平均提高系统吞吐率3.7%,对于某些特定程序组达8.5%.On shared-memory mult the main memory becomes the major iprocessors, the interconnection between the processors and bottleneck. The bandwidth-aware job scheduling is an effective way to relieve the bandwidth contention, but it requires a priori knowledge about the execution characterization of each program, which puts extra burden on system users. Moreover, the accuracy of memory bandwidth prediction is also decreased due to bandwidth contention, which limits the effectiveness of the scheduling policy. In this paper, we propose a low-overhead and transparent framework to deal with the problems. We monitor the phase changes of programs at runtime and estimate their run-alone performances. The information is updated to a database, and it can be used in the programs~ further runs to guide their scheduling, and its accuracy also evolves continuously. Evaluation results show that, with the help of the framework, the band- width aware scheduling policy can improve the system throughput by an average of 3.7 % over the native OS scheduler and up to 8.5 % improvement has been observed.

关 键 词：进程调度 访存带宽 总线竞争 跨执行优化 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]