Evaluation of HPC Applications on Cloud Resources
Abstract
We will explore the performance of traditional HPC applications and benchmarks on Cloud platforms vs. HPC specific infrastructure. We will evaluate a series of synthetic benchmarks related to flop, memory bandwidth, and network performance for MPI, as well as popular HPC applications. Comparing FutureGrid resource performance to comparable Xsede and local resources we hope to derive recommendations for to researchers which platforms should be approached based on application need.
Intellectual Merit
Understand and evaluate resources from a national resource portfolio (FutureGrid, Xsede, AWS), to build an understanding of which resources provide the most impact to given HPC research needs withing a reasonable economic model.
Broader Impact
The communities better understanding of where Cloud resources fit into a national resource portfolio for users of HPC applications should facilitate better resource matching with computational need speeding discovery and enable higher researcher productivity.
Use of FutureGrid
FutureGrid will work as the base comparison for the impact of running in a Cloud/Visualized environment.
Scale Of Use
A few VM's on unique nodes (for network effects) for a few hours over several days.