PEMPIs: A New Methodology for Modeling and Prediction of MPI Programs Performance

Abstract

The evaluation and prediction of parallel programs performance are becoming more and more important, so that they require appropriate techniques to identify the factors which influence the application execution time and also the way they interact. In this paper, we present some contributions of our research in this area by describing PEMPIs, a new methodology applied to the performance analysis and prediction of MPI programs. A new task graph helps us both to understand details of the application and to increase the accuracy of the prediction models. The proposed techniques are detailed and tested through the modeling of a complete application. PEMPIs efficiency has been proved by the results of this application modeling-most tests executed in a cluster of computers showed errors up to 10%.

Avaliando e Comparando Diferentes Estruturas de Programas Paralelos Através de Modelos Analíticos de Desempenho

“…Em trabalhos anteriores formalizamos um conjunto de estratégias para a modelagem e predic ¸ão de desempenho de aplicac ¸ões MPI. A especificac ¸ão da metodologia PEMPIs [19,18] foi desenvolvida para permitir análises de desempenho em ambientes homogêneos. Nossas avaliac ¸ões também estavam restritas a modelagem e predic ¸ão de desempenho de uma aplicac ¸ão particular.…”

Assessing contention effects of all‐to‐all communications on clusters and grids

“…To demonstrate our approach, we present experimental results obtained with different network architectures (Fast Ethernet, Gigabit Ethernet and Myrinet). We believe that this model can be extremely helpful on the development of application performance prediction frameworks such as PEMPIs (Midorikawa et al, 2004) and task schedulers.…”

Assessing Contention Effects on MPI_Alltoall Communications

“…To demonstrate our approach, we present experimental results obtained with different network architectures (Fast Ethernet, Gigabit Ethernet and Myrinet). We believe that this model can be extremely helpful on the development of application performance prediction frameworks such as PEMPIs [2], but also in the optimization of grid-aware collective communications (e.g. : LaPIe [3] and MagPIe [4]).…”