From: Norman Geist (norman.geist_at_uni-greifswald.de)
Date: Tue Apr 05 2016 - 07:09:22 CDT
Ignoring the fact that the shared PCIE bus will cause a smaller performance drop anyway, your significant slowdown may be caused by using the same cpus for your both jobs. Show us the output of
cat /proc/cpuinfo
You need to make sure each job will only use its subset of cpus. You can accomplish that using taskset, depending on how many cores you have. An example for a machine which would have 16 cores:
taskset -c 0-7 namd2 +p8 A.conf
taskset -c 8-15 namd2 +p8 B.conf
Norman Geist
> -----Ursprüngliche Nachricht-----
> Von: owner-namd-l_at_ks.uiuc.edu [mailto:owner-namd-l_at_ks.uiuc.edu] Im
> Auftrag von nikolaev_at_spbau.ru
> Gesendet: Dienstag, 5. April 2016 13:35
> An: namd-l_at_ks.uiuc.edu
> Betreff: namd-l: problem with +devices using
>
> Hello,
>
> I use a CUDA build of NAMD, and want to run 2 jobs simultaneously.
> We have 4 devices, so I assign +devices 0,1 for the first job and
> +devices 2,3 for the second.
> When I run the first job, everything is ok, and it runs well on 2 devices
> and uses 65-70% of the devices' resources.
> But when I add the second job, the performance drops immediately to 10-
> 11%
> on all 4 devices!
>
> How to solve this problem?
>
> Here is the command I run namd:
>
> /home/nikolaev_d/NAMD_2.11_Linux-x86_64-multicore-CUDA/namd2
> +setcpuaffinity +p18 +devices 0,1 +idlepoll A.conf > A.log
>
> Thank you in advance,
> Dmitrii.
>
This archive was generated by hypermail 2.1.6 : Tue Dec 27 2016 - 23:21:58 CST