[Wien] lapw2para very slow

[Peter Blaha]

Firstly, you have to find out whether it is a NFS problem as I would 
expect. Check the "timing-lines" in case.dayfile carefully. It the 
cpu-time still "normal", but the elapsed time very big (and the percentage 
rate given near the end of the line very low ?
Or is the problem in "sumpara", a final (short) step which adds up the 
parallel lapw2 results ?
Discuss with your sysadmin. On modern machines with 1Gb network, NFS 
should be fast enough. Maybe you need more NFS daemons or a operating 
system upgrade, since some Linux versions have quite slow/buggy NFS
(see the "NFS-bug" discussion on the mailing list).

Define a SCRATCH variable on the local nodes (eg. /tmp ), which of course 
should NOT be a NFS directory. This will write 
the vector and help files onto this scratch directory and thus reduce the 
NFS traffic. However, for this case you have to make sure that the number 
of k-points and processors are commensurable, i.e. for 20 k-points you can 
only use 1,2,4,5 or 10 nodes,....
 
> I'm running WIEn2k on a 16 node beowulf cluster with two CPUs per node. 
> All nodes have a common NFS.
> When I run several parallel jobs (about 3 or 4 jobs at a time using 
> 4 nodes per job), lapw1 scales almost linearly in parallel 
> but parallel lapw2 ( and sometimes lapwso) is way slower, compared to  
> a single node.
>  
>  For example, lapw2 takes about 45 seconds on 
> a singe node but for 4 nodes it can take as long as 15 minutes. 
>  
> I should state that I submit the jobs via PBS so all processes 
> for a given job (lapw0, lapw1, lapw2 e.t.c.) are confined to
> a particular set of nodes so jobs do not interfere.
>  
>  
> I find this quite confusing.
> I've run several cases and this fact is 
> consistently true. Is there any way that I can constrain lapw2 
> to run in serial mode whiles lapw1 runs in parallel?
>  
> Thanks
>  
> R. Atta-Fynn
>  
>  
> 


                                      P.Blaha
--------------------------------------------------------------------------
Peter BLAHA, Inst.f. Materials Chemistry, TU Vienna, A-1060 Vienna
Phone: +43-1-58801-15671             FAX: +43-1-58801-15698
Email: blaha at theochem.tuwien.ac.at    WWW: http://info.tuwien.ac.at/theochem/
--------------------------------------------------------------------------