利用访存负载平衡提高多核处理器系统吞吐量

Improving Multi-core System Throughput with Memory Load Balance

多任务并行执行是发挥多核处理器效能的常用手段,对多任务进行有效调度从而避免资源使用瓶颈是多核处理器任务调度的一个重要课题.本文针对多核处理器片上存储资源的使用提出一个访存负载平衡调度机制,通过刻画任务访存负载、操作系统内核层性能检测,实现了集成的计算和访存负载协同平衡系统.该调度系统在区分任务访存特征的基础上,可以在操作系统内核层组织任务执行.通过对测试程序集合的实验验证,并与其它两种调度算法进行比较,相对于操作系统缺省的计算负载平衡机制,访存负载平衡调度系统可以提高多核处理器任务吞吐量达8.2%,并且具有较低的系统实现开销.

Multi-programmed execution could exhibit higher performance with regard to single-threaded execution on multi-core plat- form, which brings multi-programmed execution into common scenario. How to schedule multiple tasks effectively in order to avoid the bottleneck of on-chip resource usage is an important issue. This paper proposed a memory load balance mechanism against on-chip memory resource underutilization. We implemented an integrated computing and memory load co-balance system via characterizing memory load and kernel level performance monitoring, which could schedule tasks at kernel level. In comparison with the default load balance scheduler, experimental results showed that memory load balance scheduling system could improve multi-core system through- put at 8.2%, and have negligible costs.