Preliminary gcc-5.0 build performance numbers

Matthew Dillon dillon at apollo.backplane.com
Wed Feb 11 16:22:42 PST 2015


    This is on an 8-thread blade.  Raw numbers:

blade (e3 8-threads):  building release and master with gcc-4.7
    release     771.73 real      3145.34 user      1223.97 sys
    master      906.42 real      4056.47 user      1394.64 sys

blade (e3 8-threads):  building master with gcc-5.0 (compare against 906.42) 
    master      964.79 real      4159.22 user      1567.45 sys without diag
    master     1258.01 real      6133.25 user      1764.31 sys with diag

--

    Crunching the numbers, remove gcc-4.4 and adding gcc-5.0 to the build,
    building world with the default compiler which is gcc-4.7, adds
    approximately 1082 cpu seconds to the build in total (user+sys), costing
    us 135 seconds in real time on the blade under test.  Considering 5.0's
    larger footprint I consider this reasonable.

    If I then set WORLD_CCVER to gcc50, to build the world using gcc-5.0
    itself, we get 965 seconds vs 906 seconds or around 60 seconds longer
    build times building with gcc-5.0 instead of gcc-4.7, which is also
    reasonable.

    However, the current gcc-5.0 in the tree is compiled with major
    diagnostics enabled and with these diagnostics the buildworld runs in
    1258 real and a hefty 2448 seconds of additional cpu time vs without the
    diagnostics.

    Since 5.0 is experimental John is considering whether to leave diagnostics
    in or not.  We want to make it easy for developers to test with gcc-5.0
    and it might be a bit too costly for people with less powerful machines.
    Of course if 5.0 were to become the default compiler they would be turned
    off.

    --

    In anycase, since I spent the day running the numbers I figured I would
    publish them so we don't lose them.

						-Matt



More information about the Users mailing list