Assembly micro-benchmark generator for characterizing Floating Point Units
Résumé
Making the right platform choice has always been a challenge for the HPC users no matter the applications vertical they are in. The number of references is very large and making the wrong choice can have adverse effects. Formerly users only had to choose between, for example, the different processors and interconnect vendors. Lately, due to the new Intel Skylake processors the choice has become increasingly difficult as different levels of performance are available within the same vendor platforms. To facilitate selection and give possible directions for the real benchmarked applications we introduce the Kernel Generator, an open source tool generating assembly kernels to help the programmer or the benchmarker understand the behavior of the different micro-architectures. We used our tool to study the behavior of the current micro-architectures and compare it to the current synthetic benchmarks which sometimes are not correctly characterizing a platform nor expose its strengths. The Kernel Generator facilitates the discovery of the platforms performance fit. To insure the relevance of our kernel, we are looking at Ansys Fluent behavior to explain the performance on the different Intel processors. In this case, we have that 4100 and 6100 Intel processors families can have equivalent performance on codes not well vectorized: Fluent being one of them. This demonstrates that we can use our tool for initial profiling and understanding of the different platforms.
Fichier principal
Jean Pourroy - Assembly micro-benchmark generator for characterizing Floating Point Units.pdf (304.29 Ko)
Télécharger le fichier
Origine | Fichiers produits par l'(les) auteur(s) |
---|
Loading...