Performance Prediction Models For Large-scale Parallel Applications

Abstract

Performance modeling plays a fundamental role in the design of computer systems. This applies especially to parallel systems where high performance is of key interest. While performance modeling of sequential computer systems already poses a number of important problems, the problem involved with performance modeling of parallel systems is even morefundamental. There exists a wide variety of approaches to the performance modeling of parallel systems, each representing a different trade-off between the accuracy of the analysis and the computational cost involved. In this diploma thesis we propose a model that attempts topredict the resources needed for the maximization of the performance of a parallel application. Similar to some of the existing techniques, the approach is primarily aimed to support the initial phases in the design of parallel systems where the emphasis is on extremely lowsolution cost, rather than on high accuracy. Our execution platform is a cluster of SMP nodes and there is an available commodity interconnect, Gigabit Ethernet. We evaluate our model in memory bound computational kernel Heat Equation in a 2 dimensional and 3 dimensionalspace, and receive accurate prediction results not only for the speedup of the applications, but the efficiency as well.