Predicting the communication pattern evolution for scalability analysis

Abstract

The performance of the message-passing applications on a parallel system can vary and cause ine ciencies as the applications grow. With the aim of providing scalability behavior information of these applications on a speci c system, we propose a methodology that allows to analyze and predict the application behavior in a bounded time and using a limited number of resources. The proposed methodology is based on the fact that most scienti c applications have been developed using speci c communicational and computational patterns, which have certain behavior rules. As the number of application processes increases, these patterns change their behavior following speci c rules, being functionally constants. Our methodology is focused on characterizing these patterns to nd its general behavior rules, in order to build a logical application trace to predict its performance. The methodology uses the PAS2P tool to obtain the application behavior information, that allow us to analyze quickly a set of relevant phases covering approximately 95% of the total application. In this paper, we present the entire methodology while the experimental validation, that has been validated for the NAS benchmarks, is focused on characterizing the communication pattern for each phase and to model its general behavior rules to predict the pattern as the number of processes increases.