new stream result

From: Greg Lindahl (glindahl@hpti.com)
Date: Mon Apr 24 2000 - 10:11:44 CDT

  • Next message: John McCalpin: "PowerComputing PowerCurve 601/120"

    If you don't have a Compaq DS10L (slate) entry yet:

    Function Rate (MB/s) RMS time Min time Max time
    Copy: 755.3222 0.0425 0.0424 0.0437
    Scale: 740.1767 0.0433 0.0432 0.0436
    Add: 661.1483 0.0726 0.0726 0.0728
    Triad: 677.7943 0.0708 0.0708 0.0709

    It only uses half of the memory slots so it's a bit slower than the DS10.
    This was

    fort -fast -tune ev6 -arch ev6

    Looks like the DS40 only has about 2 GB/s total for the 4 processors. The
    results for it are strange; I would hazard to guess that the motherboard
    chipset behaves differently under heavy load than it does under light load.
    If I run 1-3 "long" streams and then a short one, I get the same answer
    (~550 mb/s) no matter how many "longs" are runing.

    I think I should write mpistream. It will be useful for both SMP machines
    and for using stream to find out if a cluster of machines is uniform or not.
    And it's better methodology than this long/short thing.

    -- g



    This archive was generated by hypermail 2b29 : Tue Apr 25 2000 - 01:49:24 CDT