Abstract:
The resurgence of long-vector architectures (Cray XTE, NEC, EarthSimulator) in scientific computing requires different programming and modeling techniques than massively parallel processors need. Data layouts and arrangement of nested loops for high performance need to be changed, and unlike 20 years ago algorithm creators and programmers now need to deal with parallelism at the same time as vectorization. This talk will dust off some of the lessons learned 20 years ago, cover the basics of vector architectures, and provide a basic modeling methodology for use on them.