COMMTEST IBM p690 Point-to-Point Communication Performance

Performance Studies using

COMMTEST

IBM p690 SWAP Performance

(unordered swap using MPI within a node)

(performance measured per processor when all processors in node participating in a ring shift)

The following figures are concatenations of the estimated bandwidth graphs for the 8KB, 128KB, and 2MB experiments. While these experiments were not designed for determining achieveable bandwidth for message sizes other than 8 KBytes, 128 KBytes, or 8 MBytes, they are no more ill-designed for this purpose than most pingpong and pingping tests. These figures present one measure of the accuracy of these estimates. If the bandwidths for the same message sizes agree between the different experiments, it increases our faith in the validity of the estimates. Where they do not agree, the interpretation of which figure is most accurate depends on whether a single or small number of repetitions is likely to be more representative of what would be seen in practice than a larger number of repetitions. Note that measurement of a single repetition is more difficult to measure accurately, at least for the small message sizes, and may suffer more variability in timings. All of these experiments were repeated three times, however, and only the minimum times are presented.

unordered swap using nonblocking send