An Evaluation of the Performance and Scalability of a Yellowstone Test-System in Five Benchmarks

08/02/2012 - 6:00pm
FL-1001, Small Seminar Room
Ryan Pettus


The purpose of this project is to test the performance of computers with a similar hardware setup to Yellowstone, NCAR's upcoming supercomputer, as well as to provide several preliminary benchmark results. Eribus and Yakkydoodle are two test systems set up in preparation for Yellowstone, will be used in the benchmarking.  Although results of some well-known hardware benchmarks will be published upon completion in October of 2012(if the build goes as currently scheduled), some other meteorology-based benchmarks will be also be used. Two well-known benchmarks measured to measure the performance of High-Performance Computing (HPC) systems are Intel MPI Benchmark and HPC Challenge Benchmark (Sani et. Al 2008.) Some other benchmarks which replicate similar conditions to those used in atmospheric-science based operations include matrix multiplication, Fast-Fourier Transforms, and the WRF benchmark. Different physical situations will be taken into account, and the scalability of the benchmarks will be evaluated.   Additional goals will be to diagnose bottlenecks which can occur in particular when only one node is used, and compare the system to other HPC systems currently in operation.