If there is an aspect of the Linpack computation that makes it "easier" than other parallel benchmarks, it is *regularity*. The Linpack computation is fine-grained and requires much communication, but the computation is extremely *regular*, and that regularity can be exploited to minimize synchronization delays and thus keep the processors performing useful work throughout the bulk of the computation, as long as the communication requirements do not create a delay.
Link http://www.hpcwire.com/hpcwire/hpcwireWWW/04/1008/108518.html
No comments:
Post a Comment