Speaker
Peter Boyle
(University of Edinburgh)
Description
I will survey recent and emerging HPC technologies appearing in the marketplace, such as GPU’s and CPU's, and highlight the challenge of targeting all of these efficiently with a single program or simulation code. Some consideration will be given to the single node performance, multi-mode scaling performance and the use of modern 100Gbit/s+ interconnects. The effects of variable node performance (dynamic frequency scaling, Linux VM) on the collective throughput of large numbers of nodes will also be discussed.
Author
Peter Boyle
(University of Edinburgh)