Adaptive Migration Gang Scheduling for Massively Parallel Processing Systems

Abstract

As a combination of time-sharing and space-sharing, tasks are grouped together, and they are executed on processors simultaneously. Although some processors are idle, gangs are inevitably queued; therefore, gang scheduling makes fragmentation in the process space. One solution to overcome fragmentation is using migration. The benefits of migration are affected by the migration overhead. The calculation of gang scheduling is done based on Ousterhout Matrix with four phases. The second phase, called the Compact Matrix, plays a key role in gang scheduling. Improved Compact Matrix utilizes two distinct concepts, namely Difference Criterion and Exchange Idea, to improve the performance of Compact Matrix. However, with the increase in the migration overhead, the mean response time also increases. In this paper, high migration overheads are effectively prevented by calculating average migration overhead and adapting the threshold level for migration cost. The performance of the Adaptive Migration algorithm was compared to the basic Compact Matrix by simulation.