NAS Integer sort on multi-threaded shared memory machines

Abstract

Multi-threaded shared memory machines, like the commercial Tera MTA or the experimental SB-PRAM, have an extremely good performance on the Integer Sort benchmark of the NAS Parallel Benchmark Suite and are expected to scale. The number of CPU cycles is an order of magnitude lower than the numbers reported of general purpose distributed memory or shared memory machines; even vector computers are slower. The reasons for this behavior are investigated. It turns out that both machines can take advantage of a fetch-and-add operation and that due to multi-threading no time is lost waiting for memory accesses to complete. Except for non-scalable vector computers, the Cray T3E, which supports fetch-and-add but not multi-threading, is the only parallel computer that could challenge these machines.

This work was partly supported by the German Science Foundation (DFG) under contract SFB 124, TP D4.

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.