<html>
Lynne,<br><br>
<blockquote type=cite class=cite cite>Did you try recompiling lam-mpi
with the latency optimisations that were suggested (?by David) some time
ago on the list?</blockquote><br>
I certainly did. I went back, recompiled to lam-6.5.6 using
-with-tcp-short=524288 versus 64k default, --with-rpi=usysv<br><br>
For those interested, here are the details for the system along with the
scaling numbers and details:<br><br>
System:<br>
<x-tab> </x-tab>Intel P733
Dual Processor<br>
<x-tab> </x-tab>100 base
Ethernet<br>
<x-tab> </x-tab>Linux
Redhat 7.3<br>
<x-tab> </x-tab>lam-6.5.6
(tcp-short=524288)<br>
<x-tab> </x-tab>fftw-2.1.3
(enable-mpi)<br>
<x-tab> </x-tab>gromacs-3.1.4
(enable-mpi)<br><br>
MD Simulation:<br>
<x-tab> </x-tab>7.5 nm
cube<br>
<x-tab> </x-tab>16
molecules, 16 Na, + water<br>
<x-tab> </x-tab>42,000
atoms<br>
<x-tab> </x-tab>500
steps<br>
<x-tab> </x-tab>0.002 ps
step<br>
<x-tab> </x-tab>PME
(order=4, 0.12nm fourier spacing)<br><br>
grompp switches:<br>
<x-tab> </x-tab>-sort<br>
<x-tab> </x-tab>-shuffle<br>
<x-tab> </x-tab>-np
#<br><br>
Initiation of mdrun:<br>
<x-tab> </x-tab>mpirun C
mdrun -s .....<br>
<x-tab> </x-tab>1 CPU run
time approximately 500 sec<br><br>
SCALING<br>
Cluster<x-tab> </x-tab>tcp=64k<x-tab> </x-tab>tcp=524k<x-tab> </x-tab>PME
off<x-tab> </x-tab><x-tab> </x-tab>PME(6/0.17nm)<br><br>
1 box/2
CPU<x-tab> </x-tab>71%<x-tab> </x-tab><x-tab> </x-tab>72%<x-tab> </x-tab><x-tab> </x-tab>86%<x-tab> </x-tab><x-tab> </x-tab>84%<br>
2 box/2
CPU<x-tab> </x-tab>54%<x-tab> </x-tab><x-tab> </x-tab>56%<x-tab> </x-tab><x-tab> </x-tab>83%<x-tab> </x-tab><x-tab> </x-tab>74%<br>
2 box/3
CPU<x-tab> </x-tab>40%<x-tab> </x-tab><x-tab> </x-tab>41%<x-tab> </x-tab><x-tab> </x-tab>73%<x-tab> </x-tab><x-tab> </x-tab>62%<br>
2 box/4
CPU<x-tab> </x-tab>27%<x-tab> </x-tab><x-tab> </x-tab>30%<x-tab> </x-tab><x-tab> </x-tab>71%<x-tab> </x-tab><x-tab> </x-tab>51%<br><br>
So:<br>
<x-tab> </x-tab># changing
the tcp-short of lam did improve things slightly, but not really that
much to justify the right royal pain it was ;-)<br>
<x-tab> </x-tab># PME
doesn't scale very well, as is noted in the manual etc.<br>
<x-tab> </x-tab># PME does
scale better when the fourier grid spacing is increased and the PME order
increased<br><br>
I am going to see if there is much of a difference for a longer run, to
reduce the effect of the setup and shut down period that doesn't
parallelise, as noted by Anton.<br><br>
Catch ya,<br>
<x-sigsep><p></x-sigsep>
<b>Dr. Dallas Warren<br>
</b><i>Research Fellow<br>
</i>Department of Pharmaceutical Biology and Pharmacology<br>
Victorian College of Pharmacy<br>
Monash University<br>
<font color="#0000FF"><u>dallas.warren@vcp.monash.edu.au<br>
</u></font>+61 3 9903 9076</html>