[torqueusers] Code seems to run slow when run with pbs script

James J Coyle jjc at iastate.edu
Thu Nov 30 09:15:35 MST 2006


Nick,

  When the job is running, login to the first node listed for your job 
when you use qstat -n node (if you can) and use top.  
My guess is that you'll see all your MPI processing running on the 
first node.

  Tourque will assign you nodes to use, but you have to tell MPI
that you want to use the nodes.

  My guess is hat you have not included  

   -machinefile $PBS_NODEFILE

on the mpirun command.

  If not, then you are using the default created by PGI, and may very 
well just be localhost.  This means when you run with -np 2 through 64 they
all just run on the single node, with no speedup.

  My guess is that all your cluster nodes are single CPU so you get 14 sec. 
per iteration, and when this same thing happens on your dual CPU host, 
you see the 2x performance improvement due to the 2 processors.


  Other (less likely) possibilities are that there are orphaned processes
from previous MPI runs or some other user is running on your nodes by
using a machinefile other than his/her own PBS_NODEFILE.

-- 
 James Coyle, PhD
 SGI Origin, Alpha, Xeon and Opteron Cluster Manager
 High Performance Computing Group     
 235 Durham Center            
 Iowa State Univ.           phone: (515)-294-2099
 Ames, Iowa 50011           web: http://jjc.public.iastate.edu

> Ive got an issue that is odd.
> When we submit a job inside of a pbs script it seems to run pretty slow.
> For example a single iteration of this code takes about 7 seconds on a dual 
> cpu opteron at 2.6Ghz.
> When running on a Cluster with 2.4Ghz cpus inside of a pbs script a single 
> iteration takes approximatly 14 seconds. It dosnt matter how many nodes the 
> job is run on. 2 - 64 gets the same times per iteration.
> 
> But if we run manually with mpirun it takes about 1 tenth of a second.
> 
> Any ideas?
> 
> _______________________________________________
> torqueusers mailing list
> torqueusers at supercluster.org
> http://www.supercluster.org/mailman/listinfo/torqueusers
> 




More information about the torqueusers mailing list