multicore vs ibverbs

From: Gianluca Interlandi (gianluca_at_u.washington.edu)
Date: Sat Jul 04 2015 - 12:52:40 CDT

Dear NAMD list,

I have been comparing the multicore version of NAMD 2.9 vs ibverbs on a
single node of the SDSC comet cluster. A node consists of 2 processors
with 12 cores each (24 cores in total). I have compiled both versions
locally on comet.

I am able to run the multicore version using "namd2 +p24" (without
charmrun) and I noticed that it is ca. 5-7% faster than using the ibverbs
version, probably because it lacks the overhead of the network libraries.

Just wondering: is there any effect of the network libraries onto the
"accuracy" of NAMD? Or, is the result somehow influenced by the choice of
network layer (or lack thereof)? I mean, if I compare multicore vs ibverbs
while keeping everything else identical (no. of cores, exact same node,
same random seed, same load on each core, etc.), is the trajectory still
expected to be different? Lastly, is the domain decomposition affected by
the choice of network libraries?

Thanks a lot in advance!

                       Gianluca

-----------------------------------------------------
Gianluca Interlandi, PhD gianluca_at_u.washington.edu
                     +1 (206) 685 4435
                     http://artemide.bioeng.washington.edu/

Research Assistant Professor at the Department of Bioengineering
at the University of Washington, Seattle WA U.S.A.
-----------------------------------------------------

This archive was generated by hypermail 2.1.6 : Thu Dec 31 2015 - 23:21:57 CST