From: Christopher Hartshorn (cmhartshorn_at_wsu.edu)
Date: Mon Nov 24 2008 - 17:03:40 CST
Hello everyone. I have a few questions as to optimizing NAMD to
current MacPro with dual 4-core Intel Xeons. First, off even though
there are 8 cores total in these machines, when I run NAMD with:
charmrun ++local namd2 +p8
I get much slower (almost half speed) performance than if use a +p4
option (also if I use +p2 the speed is cut in half).
Next, is there any gains to be had from playing with charmrun options
like "++p n" or ++pe for the MacPro.
Next, in the optimization page I see that numprocs=numpatch+1 scaled
to number of cores. If I have a 5x5x5 patch grid (125=npatches) and I
have 8cores would the optimized value be 128 because 125+1=126 then
scaled to the next multiple of 8. And where is this value set? Is
this +p128? Or is this put into one of the ++p options in charmrun?
Finally, I noticed that I can not set the +setcpuaffinity option for
charmrun which got me thinking that I do not have the latest charmrun
compiled. I have simply downloaded the binaries from the NAMD website
which were compiled over two years ago for 32bit Intels. I have the
latest Intel compilers, so would it be wise just to get the latest
source for Charmm++ and the bleeding-edge NAMD and compile them myself.
What got me into this idea that I needed to optimize in the first
place is that I installed the binaries on the above computer as well
as on our cluster here at WSU. I get the following results between
the two computers on the exact same initial MDSim:
2nodes of cluster (each node has 2x2core AMD 64bit opterons running at
2.4GHz) total of 8cores=14min/ps
MacPro (main motherboard has 2x4core Intel 64bit Xeons running at
2.83GHz) total of 8cores=31min/ps
Thus, the MacPro is considerable slower than the same number of slower
Opterons. Only difference would be compilers of the respective
binaries and some OSX overhead on the MacPro (which could not possibly
account for the massive difference between the two). Any help and/or
thoughts on this would be greatly appreciated. Thank you in advance.
Christopher M Hartshorn
This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:50:09 CST