Selection of message passing collectives in presence of system noise
Consideration of system jitter in selecting a “message passing collectives algorithm” used in a message passing interface. A set of multiple message passing collectives algorithms are ranked against each other with at least some consideration of relative jitter-related performance as between the algorithms. The consideration of jitter includes consideration of “system jitter” (OS jitter and/or network jitter). In some embodiments, multiple rankings are performed for different levels of system jitter.
1. A computer-implemented method comprising:
providing a plurality of message passing collectives algorithms;
evaluating each algorithm of the plurality of message passing collectives algorithms for performance when subject to at least one type(s) of jitter to obtain jitter evaluations;
ranking the plurality of message passing collectives algorithms relative to each other based upon the jitter evaluations to obtain a first ranking;
selecting, at runtime, a first message passing collectives algorithm from the plurality of message passing collectives algorithm, where the selection is based, at least in part, on the first ranking;
performing message passing using the first message passing collectives algorithm; and
estimating a level of system jitter that will occur at runtime through performance data collected by benchmarks during offline calibration to obtain an estimated system jitter level;
wherein:
the selection of the first message passing collectives algorithms is further based, at least in part, upon a first non-jitter factor which is not related to any type of jitter; and
the at least one type(s) of jitter includes both of the following types of jitter: operating system jitter and/or network jitter.