Performance Variability of Highly Parallel Architectures

摘要

The design and evaluation of high performance computers has concentrated on increasing computational speed for applications. This performance is often measured on a well configured dedicated system to show the best case. In the real environment, resources are not always dedicated to a single task, and systems run tasks that may influence each other, so run times vary, sometimes to an unreasonably large extent. This paper explores the amount of variation seen across four large distributed memory systems in a systematic manner. It then analyzes the causes for the variations seen and discusses what can be done to decrease the variation without impacting performance.