There goes the neighborhood

Abhinav Bhatelé, Kathryn Mohror, S. H. Langer, Katherine E. Isaacs

2013 · 170 citations · 13 references

DOIFull text

Open access

Concepts

TL;DR

Predictable performance is essential for diagnosing application issues, assessing code or system changes, estimating batch job runtimes, and allocating resources. The study investigates causes of performance variability on Cray machines, including OS jitter, partition shape, and network interference. The authors analyze run‑to‑run variability on Cray XE systems by measuring execution times of a communication‑heavy parallel application under different OS, partition, and network conditions. Experiments on Cray XE systems show that the application’s execution time can vary from 28 % faster to 41 % slower, while Blue Gene systems exhibit no variability, suggesting that reducing such fluctuations could improve throughput and lower energy costs.

Abstract

Predictable performance is important for understanding and alleviating application performance issues; quantifying the effects of source code, compiler, or system software changes; estimating the time required for batch jobs; and determining the allocation requests for proposals. Our experiments show that on a Cray XE system, the execution time of a communication-heavy parallel application ranges from 28% faster to 41% slower than the average observed performance. Blue Gene systems, on the other hand, demonstrate no noticeable run-to-run variability. In this paper, we focus on Cray machines and investigate potential causes for performance variability such as OS jitter, shape of the allocated partition, and interference from other jobs sharing the same network links. Reducing such variability could improve overall throughput at a computer center and save energy costs.

References

13