On the problem of minimizing workload execution time in SMT processors

Abstract

Most research work on (simultaneous multithreading processors) SMTs focuses on improving throughput and/or fairness, or on prioritizing some threads over others in a workload. In this paper, we discuss a new problem not previously addressed in the SMT literature. We call this problem workload execution time (WET) minimization. It consists of reducing the total execution time of all threads in a workload. This problem arises in parallel applications, where it is common for a single master thread to spawn several child jobs. The master job cannot continue until all child jobs have finished. Reducing the overall execution time is important to speedup the application. This paper is a first step in analyzing this problem. First, we analyze the WET provided by the best fetch policies turned at improving throughput/fairness. We demonstrate that these policies achieve less than optimum performance. We show that, on average, for the workloads evaluated in this paper, there is space for improvement of up to 18 percentage points. It follows that novel mechanisms trying to reduce WET are required to speedup parallel applications.