[mpich-discuss] MPICH on TILE-Gx

[Bryant Lam]

I'm currently working on some performance numbers of MPICH 3.1.3 on the 
Tilera TILE-Gx platform. While direct support isn't indicated, I wanted 
to see what the performance profile would be for some of the core MPI 
functions. The barrier times, however, exhibit poor performance and high 
variance as the number of PEs scale up on this 36-core SMP device.

    PEs    Trial 1     Trial 2(times in microseconds)
    2         7.28        7.54
    3        11.63    13010.92
    4        15.01      268
    5     21073.14       18.11
    6        19.63       18.97
    7        17.83    22698.7
    8        23.82     4099.2
    9        24.72     9334.79
    10    39859.54     1591.71
    11    39586.55    38599.86
    12    30229.56    39799.5
    13    28860.67     5684.71
    14    38742.38    39956.55
    15    43117.49    39952.83
    16    26762.15    26230.78
    17    39458.78    59911.38
    18    46232.86    45327.22
    19    45462.67    54762.76
    20    59744.57    39657.49
    21    54023.2     72209.07
    22    78213.27    49108.57
    23    69221.4     79534.28
    24    71599.46    81920.28
    25    45443.61    76059.46
    26    50649.56    72402.17
    27    39936.37    66906.83
    28    64324.64    44803.16
    29    59913.22    53071.9
    30    44569.65    70844.15
    31    55428.5     47399.51
    32    54708.82    64102.45
    33    34413.99    52384.35
    34    66390.67    97134.79
    35    68839.46    62719.52
    36    43243.98    59132.83(59.1 ms)

Each entry is an average of 100 iterations. Benchmark used is from OSU 
MPI microbenchmarks (osu_barrier.c).

These are the configure options I used to cross-compile and build MPICH 
for TILE-Gx:

    mkdir build && cd build
    PREFIXROOT=/opt/tilegx
    ../configure CC=tile-gcc CXX=tile-g++ FC=tile-gfortran
    F77=tile-gfortran \
         --prefix=$PREFIXROOT/mpich --host=tile \
         --disable-fortran \
         --enable-fast=yes --with-thread-package=posix


Are there any build options that I'm not using that I should be? I'm 
trying to focus this build specifically for SMP performance.

Thanks,

Bryant Lam





-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpich.org/pipermail/discuss/attachments/20150202/c9973e0d/attachment.html>
-------------- next part --------------
_______________________________________________
discuss mailing list     discuss at mpich.org
To manage subscription options or unsubscribe:
https://lists.mpich.org/mailman/listinfo/discuss