Understanding Application Scaling: NAS Parallel Benchmarks on the NOW and SGI Origin 2000
We present a study of the architectural requirements and scalability of
the NAS Parallel Benchmarks. We find that the local processor and memory
systems dominate both the overall performance and scalability of the NAS
parallel benchmarks, as opposed to the communication architecture.
Improvements in computational efficiency under scaling, due to the
increase in global cache size, often offset the rising costs of
communication and synchronization, hence yielding linear or even
super-linear speedups. We also find that communication performance is
often much lower than would be expected from micro-benchmark results.
For more information, see
this paper.
Presentation: PowerPoint slides,
postscript gzipped,
postscript 2up gzipped.
Click here to
go back to the Finale schedule.