Benchmark based on application signature to analyze and predict their behavior

Abstract

Currently, there are benchmark sets that measure the performance of HPC systems under specific computing and communication properties. These benchmarks represent the kernels of applications that measure specific hardware components. If the user’s application is not represented by any benchmark, it is not possible to obtain an equivalent performance metric. In this work, we propose a benchmark based on the signature of an MPI application obtained by the PAS2P method. PAS2P creates the application signature in order to predict the execution time, which we believe will be very adjusted in relation to the execution time of the full application. The signature has two performance qualities: the bounded time to execute it (a benchmark property) and the quality of prediction. Therefore, we propose to extend the signature by giving the benchmark capacities such as the efficiency of the application over the HPC system. The performance metrics will be performed by the benchmark proposed. The experimentation validates our proposal with an average error of prediction close to 7%.