异构机群下最小化非实时任务平均响应时间的调度策略

A Job Scheduling Policy Used to Minimize the Mean Response Time of the un-Real Time Jobs in the Multi-Heterogeneous Clusters

机群作为高性能计算机平台有着广阔的应用前景。由于很多情况下单一机群已经无法满足应用需求,因此分离的机群通常相互连接来建立一个多机群计算结构。分布式系统中根据系统处理任务性质的不同可分为实时任务和非实时任务两类。它们继而又可以分为有数据关联任务和无数据关联任务。本文关注的是异构机群下非实时任务流的调度,采用最小化任务平均响应时间作为目标函数。已有的策略要么不适用于异构多机群,要么没有考虑到任务传递的通信耗费,或者两者都欠缺。本文提出的任务调度策略解决了这两个问题,并且通过试验证明了MMRT虽然充分考虑了通信耗费,却没有明显增大系统的额外开销。

Clusters are more and more popular being used as the high performance computing platform. Clusters need to connected together to form a much stronger integrated computing system to fulfill the request of the applications. The jobs which are processed by the system are divided into two main kinds, refered as real-time jobs and un-real-time jobs, and each can be further divided into data related jobs and un-data related jobs. This paper address on processing of the un-real-time jobs without data relationships. And made the goal function as the minimizing the mean response time. Some of the existed polices are not suited in the multi-heterogeneous-clusters, others are lack of communication cost considering, or both. This paper propose a job scheduling policy to solve those problem without increasing the additional system cost.