[erlang-questions] wfinder8 on the T2150

Anders Nygren anders.nygren@REDACTED
Sat Nov 3 16:13:02 CET 2007


Tim has now sent me the results of running my wfinder8 on the
T2150. Here is a summary of the data.

This is based on Sun's new Niagara T2 processor
8 cores
2 integer units/core
8 threads/core
1 4 MB L2 cache

Solaris and erlang sees this as an 64 processor machine.

schedulers  real     user    sys     total  total/real
1         36.108    31.174   4.720   35.894   0.99
2         18.884    31.576   4.826   36.403   1.93
4         10.502    33.026   5.124   38.150   3.63
8          6.394    34.943   5.494   40.437   6.32
16         4.741    38.479   7.404   45.883   9.68 (1)
32         4.420    46.277  12.267   58.544  13.25 (1)
64         4.447    56.132  18.923   75.055  16.88 (1)
128        6.018  2:20.710  24.746  165.456  27.49 (1)

(1) When there are more than 8 threads/processes in
    run state the CPU is not able to correctly track
    the actual running time.

Speedup
1-2    1.91
2-4    1.80
4-8    1.64
8-16   1.35
16-32  1.07
32-64  0.99
64-128 0.74

1-2    1.91
1-4    3.44
1-8    5.65
1-16   7.61
1-32   8.17
1-64   8.12
1-128  6.00

So we get a maximum speedup of a factor 8.

And erlang is only 0.04 seconds slower than a Python program
that uses separate OS processes to get around the Python
global interpreter lock and mmap.
And of course that 0.04 seconds is well within variation from
run to run.

/Anders



More information about the erlang-questions mailing list