A Simulation Study of Hardware Parameters for GPU-based HPC Platforms
TimeThursday, 19 November 20208:30am - 5pm EDT
DescriptionHigh-performance computing (HPC) platforms are switching to GPU-based compute nodes; the resulting trend is the increase in per-node computational capacity and the reduction of the number of endpoints in the system. This trend changes the computation and communication balance in comparison to the pre-GPU era HPC platforms, and warrants a re-study of the hardware architectural parameters. In this research, we perform a simulation study of the impact of crucial hardware parameters in GPU-based systems using HPC workloads that consist of representative HPC applications. The hardware parameters studied include link bandwidth, number of GPUs per node and interconnection network topology.