Abstract |
---|
The BSP cost model measures the cost of communication using a singlearchitectural parameter, $g$, which measures permeability of thenetwork to continuous traffic. Architectures, typically networks ofworkstations, pose particular problems for high-performancecommunication because it is hard to achieve high throughput, and evenharder to do so predictably. Yet both of these are required for BSP tobe effective. We present a technique for controlling appliedcommunication load that achieves both. Traffic is presented to thecommunication network at a rate chosen to maximise throughput andminimise its variance. Performance improvements as large as a factorof two over MPI can be achieved. |
Contact |
Stephen Donaldson Oxford University Computing Laboratory,Wolfson Building ,Parks Road,Oxford,OX1 3QD,United Kingdom Stephen.Donaldson@comlab.ox.ac.uk |