[Myrinet] Using mpd with gm apps in mpich-1.2.4..8a (performance difference)

I am trying to configure my cluster to run
using the MPD process to help speedup large
job startup times.   I installed gm-1.5.2.1
and mpich-1.2.4..8a on a couple of dual Xeon
systems (PCI64B).  Here are some results from
running with mpirun.ch_gm and mpirun.mpd (see
below).  The machfn.test file looks like:

g0128
g0129

The bandwidth is about 1/2 of what I expect and
the latency is about 50% worse when using mpd to
start the program.  Am I doing something wrong?

Thanks,
Craig

(Redhat 7.2, linux-2.4.19, mpich and mpptest were
built using the Intel compilers).


[root@g0128 mpptest]# mpirun.ch_gm -machinefile machfn.test -np 2 ../perftest/mpptest -size 0 1048576 65536
set default
set font variable
set curve window y 0.15 0.90
set order d d d x y d
title left 'time (us)', bottom 'Size (bytes)',
      top 'Comm Perf for MPI (g0128)',
 'type = blocking'

#p0     p1      dist    len     ave time (us)   rate
0       1       1       0       10.395938       0.00
0       1       1       65536   335.955000      195.074e+6
0       1       1       131072  613.970000      213.483e+6
0       1       1       196608  892.130000      220.380e+6
0       1       1       262144  1170.655000     223.929e+6
0       1       1       327680  1449.845000     226.010e+6
0       1       1       393216  1728.205000     227.529e+6
0       1       1       458752  2006.880000     228.590e+6
0       1       1       524288  2284.545000     229.493e+6
0       1       1       589824  2564.125000     230.029e+6
0       1       1       655360  2842.375000     230.568e+6
0       1       1       720896  3120.380000     231.028e+6
0       1       1       786432  3399.500000     231.338e+6
0       1       1       851968  3677.945000     231.642e+6
0       1       1       917504  3956.155000     231.918e+6
0       1       1       983040  4236.020000     232.067e+6
0       1       1       1048576 4514.125000     232.288e+6

[root@g0128 mpptest]# mpirun.mpd -y -m machfn.test -np 2 ../perftest/mpptest -size 0 1048576 65536
set default
set font variable
set curve window y 0.15 0.90
set order d d d x y d
title left 'time (us)', bottom 'Size (bytes)',
      top 'Comm Perf for MPI (g0128)',
 'type = blocking'

#p0     p1      dist    len     ave time (us)   rate
0       1       1       0       14.569375       0.00
0       1       1       65536   520.035000      126.022e+6
0       1       1       131072  994.175000      131.840e+6
0       1       1       196608  1471.060000     133.651e+6
0       1       1       262144  1948.785000     134.517e+6
0       1       1       327680  2427.970000     134.960e+6
0       1       1       393216  2905.945000     135.314e+6
0       1       1       458752  3383.055000     135.603e+6
0       1       1       524288  3862.385000     135.742e+6
0       1       1       589824  4340.945000     135.875e+6
0       1       1       655360  4817.245000     136.045e+6
0       1       1       720896  5297.255000     136.089e+6
0       1       1       786432  5774.510000     136.190e+6
0       1       1       851968  6253.060000     136.248e+6
0       1       1       917504  6731.065000     136.309e+6
0       1       1       983040  7209.530000     136.353e+6
0       1       1       1048576 7687.510000     136.400e+6
plot square
join
wait
new page


-- 
Craig Tierney ([email protected])