Enhancement and Implementation of MCL HPL Performance Modeler
Young Woo Kim · 2021
In this paper, an enhanced Multi-Communication Layered HPL (MCL HPL) model and its implementation is described Multi-Communication Layered HPL model is a numerical model to estimate the performance of HPL – a well-known de-facto benchmark for measuring a high performance computing system - with less estimation error by modeling multiple network layers for communication overhead calculations in HPL. The original MCL HPL model assumed no memory read and write overhead in each communication layers. In real operation, network packet transmission incurs memory read and write at source and target location. If the speed of network is relatively higher than the memory access, the performance of overall network data transfer will degrade due to the lower memory access speed The original MCL HPL model assumed the zero-copy without consideration of memory access in calculating communication overhead, and this amplifies the error in performance estimation. In this paper, MCL HPL model is modified to considering the memory access in calculation of communication overhead, and the enhanced MCL HPL performance modeler is implemented and tested. The proposed MCL HPL modeler is applied to estimate the performance of well-known TOP500 systems and compared the results to the original modeler and the real performance values. The experimental results showed that the enhanced HPL performance modeler reduces more errors (1~3%) in estimation compared to the MCL HPL modeler.