Performance Analysis of the MPAS-Ocean Code using HPCToolkit and MIAMI

MPAS-Ocean [4] is a component of the MPAS framework of climate models. MPAS-Ocean is an unstructured-mesh ocean model capable of using enhanced horizontal resolution in selected regions of the ocean domain. The code is publicly available for download [3] and comes with several input problems of different sizes corresponding to different simulation resolutions. In this initial study, we look at the per-core performance of version 2.0 of the MPAS-Ocean code. Our analysis was performed on a single node system with dual Intel Xeon E5-2690 CPUs, based on the Sandy Bridge micro-architecture. Each processor has 8 cores and a shared 20 MB L3 cache. We compiled the code with the Intel Fortran compiler 14.0.0 and optimization flags-O3-g. Figure 1: HPCToolkit time profile: the MPAS code spends 38.6 % of time in the compiler intrinsic ’trim’ invoked primarily from the software timer callipers. Due to the small system size, we used an input problem of lower resolution (240km), and performed runs with one and four MPI processes. Figure 1 shows an initial HPCToolkit [1] time profile of the application running with a single MPI process. We use HPCToolkit’s callers view and sort the results 1 Figure 2: HPCToolkit time profile: the top time consuming routines in MPAS-Ocean after fixing th