Análisis de rendimiento de un clúster HPC y, arquitecturas manycore y multicorer

Abstract

Currently the tendency to obtain a large amount of computing, is through parallel computing, a clear example lies in the fact that the fastest computers in the world are clusters for high performance computing, formed by accelerators and multiprocessors. HPC clusters offer an important computational capacity to solve high computational cost requirements, in areas such as: weather prediction, machine learning. However, to achieve optimal performance is necessary to analyze their requirements, such as: communications, parallelization platforms. The objective of this study is to analyze the scalability of the climate prediction model WRF in HPC clusters based on inter-node communication and MPI processes; and the performance of the Horizontal Diffusion algorithm on XeonPhi and TeslaKepler accelerators, using OpenCL and CUDA. The results show a dependence of the scalability of the WRF with communications, since; a maximum acceleration was obtained for InfiniBand FDR of 25.9, QDR 21.42 and Ethernet 6.4 times more than sequential execution. The accelerator Tesla K40m shows a performance 6 times greater than XeonPhi, since the algorithm does not efficiently use vectorization Intel property, in addition Intel OpenCL drivers for its architecture manycore, they are deprecated. OpenCL has a higher learning curve than CUDA due to its multiplatform properties, in terms of performance CUDA shows a 6% improvement since it is oriented to NVIDIA GPUs, and has configurations that improve the performance, on the other hand OpenCL is not very affected by its multiplatform property, and is an option to consider if the requirements are oriented towards portability.