Scalability analysis of large codes using synthetic perturbations and factorial designs.

Abstract

A class of interaction plots using speedup is introduced in this dissertation that will enable the investigator in comparing the scalability of two parallel systems.

Scalability will be limited by serial bottlenecks in the code. Locating these bottlenecks in parallel environment is not trivial. We used factorial designs to estimate empirically an approximation of a multivariate Taylor's expansion for the code's execution response function. The first order terms in the Taylor's expansion function correspond to the suspected bottlenecks and scale factors. The coefficients of these terms are estimates of the code's sensitivity to changes in these suspected locations and scale factors. The higher order terms are utilized as informal relative indicators of the code's scalability. This approach was applied to a large meteorology code running on the CRAY J90 and the IBM SP2 scalable distributed memory machine.

The issue of code's scalability is becoming more crucial with the existence of advanced scalable architectures. While speedup relates the reduction in time when going from serial to parallel computation, scalability focuses on the overall performance resulting from the increase in problem size and the number of processors.