Evaluating OpenMP Performance on Thousands of Cores on the Numascale Architecture

摘要

The company Numascale is one of the few companies offering shared memory systems with thousands of cores. To produce machines of such a size, a proprietary cache-coherent interconnect is used to couple standard servers into a large single system. In this work we investigate the ability of such a huge system to run OpenMP applications in an efficient way. Therefore, we use kernel benchmarks to investigate basic performance characteristics of the machine and we present a real world application from the Institute of Combustion Technology at RWTH Aachen University, called TrajSearch, which has been optimized to run efficient on such a large shared memory machine.