NAMD infiniband performance

From: Guanglei Cui (amber.mail.archive_at_gmail.com)
Date: Mon May 18 2009 - 16:10:55 CDT

Hi NAMD experts,

I have some questions about NAMD infiniband performance, and would
like to hear the expert's opinions. The IT department here is looking
for justifications on spending on infiniband. MD simulations are one
of the applications that are being considered as showcases for the
need of high performance network hardware. The test machines are 3
24-core boxes connected with Giga-E. The CPU type is Intel Xeon
E7450_at_2.40GHz. Having not spent a lot of time on it, I did some simple
benchmarks using the pre-compiled NAMD binary for x64, and didn't see
much difference in performance. I was hoping to see some impact due to
the network communication. I know I may not be getting that because of
the binary I used. The benchmark is the standard Apoa1 system, a
16-CPU job allocated in one box or split between two boxes. I guess I
probably should compile NAMD to take advantage of SMP. So, my
questions are the following,

1) what is the proper way to benchmark NAMD on the hardware platform
in order to evaluate the need for infiniband?
2) what is your recommendation on compiling NAMD using Intel Compiler
11.0 on the given hardware platform that achieves the best
performance? My locally compiled NAMD executable is still a little
slower than the precompiled one.

I apologize if similar questions have been previously discussed here
or elsewhere. You input is greatly appreciated.

Best regards,

-- 
Guanglei

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:50:59 CST