OS Scheduling for Simultaneous Multithreading Processors

by Jonathan Ledlie, Matt McCormick, and Omer Zaki

Within recent years the concept of the simultaneous multithreading (SMT) processor has been gaining in popularity. This hardware allows multiple processes to run on the processor at the same time providing more potential for instruction level parallelism. These new processors suggest that the rules an operating system (OS) scheduler follows need to be changed or at least modified. Our study shows the combination of jobs selected to run on these threads can significantly affect system performance. Our research shows that scheduling policies are greatly affected by the system workload and there most likely does not exist a single, best scheduling policy. However, it can be shown that a scheduler that tries to schedule processes doing a large number of loads and stores together with jobs doing few loads and stores consistently performs at levels close to or better than all other scheduling policies examined. It can also be seen that the more possibilities there are for scheduling, the more necessary it is to have an intelligent scheduler. In contrast, the few number of decisions to make (few threads and/or few processes) the less important the decision of a scheduler becomes.

View paper in pdf format or in ps format.