On System Noise and Large-scale Parallel Applications

Date

May 26, 2011

Time

3:00 PM - 4:00 PM

Speaker

Torsten Höfler

Affiliation

University of Illinois, USA

Series

TUD ZIH Kolloquium

Language

en

Main Topic

Informatik

Other Topics

Mathematik, Informatik

Host

Ralf Müller-Pfefferkorn

Description

System noise is increasingly a concern as HPC systems continue to grow in scale. This talk presents an in-depth analysis of the impact of system noise on large-scale parallel application performance in realistic settings. We present a model-based simulation toolchain that uses the LogGOPS model and Lamport's clock condition to inject noise delays from traces into complex communication relations. We show that the number of processes at which noise becomes a significant bottleneck is system-specific and depends more on the structure of the noise than its intensity. In contrast to previous studies, we show that not only collective operations but also point-to-point communications influence the application's sensitivity to noise. We also discuss the influence of noise-mitigation strategies, for example, our results verify the long-standing conjecture that co-scheduling prevents significant application slowdown. We discovered that the noise bottleneck can lead to a 100x slowdown of single collective operations and a 2x slowdown of scientific applications on very large-scale systems. Our developed tools can be utilized quantify this noise bottleneck and to tune the noise signatures of systems and applications for minimal noise propagation.

Last modified: May 26, 2011, 2:07:04 PM