Scalability of Parallel Applications: An approach to predict the computational behavior

摘要

When a message-passing application is executed many times over a long period of time, using an elevated number of resources, it is critical to predict its behavior before executing it. We propose a methodology to predict the strong scalability behavior for message-passing applications in specific systems. It is focused on characterizing and analyzing the communication and computational application patterns, from a set of executions in small scale, to project their behavior when the number of processes increases. The methodology strives to use a reduced number of resources. This paper presents the general methodology, focusing on validating the computational time model, which is a regression based approach. This model allows us to predict the computation time with high accuracy for a large number of processes. We executed from 16 to 256 processes and we predicted the computation time up until 4,096 processes. For the applications tested, we obtained an error of less than 9%.