From: Baker D.J. (D.J.Baker_at_soton.ac.uk)
Date: Thu Jun 25 2015 - 07:00:49 CDT
We have just acquired an IBM POWER8 machine with a K40 GPU card installed. I'm starting to do some experiments with NAMD on the machine, and I'm impressed. We're using Ubuntu 14.10, and I have built NAMD 2.10 using a recent version of the GNU compilers. (For some reason I couldn't manage to build namd2 using IBM's compilers, however charm was OK.) So, as I say, the GNU compiled binary does reasonably well on the host. For example, I have run the stmv benchmark over 4 processors and the GPU card and the simulation was completed in 185s.
I wonder if anyone has any experience of running NAMD on a POWER machine (I know that POWER8 machines aren't too common just yet), and has any tips to optimize the performance of the package on that architecture, So at the moment I'm keeping things simple, this is my command line:
charmrun +p4 namd2 +idlepoll stmv.namd >stmv.out
I'm an administrator, not a NAMD user, and so I would value your advice, please. For example, would turning on SMT be worthwhile and/or should I use processor affinity? If processer affinity is important then I would value an example, please.
This archive was generated by hypermail 2.1.6 : Thu Dec 31 2015 - 23:21:56 CST