Predicting parallel applications' performance across platforms using partial execution
Tao Yang, Xiaosong Ma, Frank Mueller · NCSU Libraries Repository (North Carolina State University Libraries) · 2005
Performance prediction across platforms is increasingly important in today's diverse computing environments.As both programs and their developers face unprecedented wide choices in execution platforms, cross-machine execution time prediction with reasonable accuracy equally benefits scheduling decisions of grid jobs as well as scientists in their research and development planning.In this paper, we investigate an affordable method approaching cross-platform performance translation, based on the notion of relative performance between two platforms.We argue that relative performance can often be observed without running a parallel application in full.This paper shows that it suffices to observe very short partial executions of an application since most parallel codes are iterative and behave in a predictable manner after a minimal startup period.This prediction approach is observationbased and does not require program modeling, code analysis, or architectural simulation.Our performance results (using four real-world parallel simulation codes and a total of ten parallel machines with eight distinct architectures) demonstrate that performance prediction derived from partial application executions can yield highly accurate results at a low cost.