[CPMD-list] slow GLOSUM subroutine

[Axel Kohlmeyer]

On Mon, 2 Jul 2007, Xevi Biarnés (PCB) wrote:


hi xevi,

XB> I am experiencing a lot of elapsed time on subroutine GLOSUM (upto
XB> 15 times CPU time)
XB> 
XB> I run CPMD version 3.9.1 on a 115 atoms system (LSD calculation) ==>
XB> 358124 PW
XB> 
XB> I run on a IBM distributed memory machine (96 processors) with a
XB> myrinet network. System administrators tell me that there are no
XB> network problems

XB> 
XB> Did anybody experience something similar?
XB> 
XB> any ideas? solutions?

no, but this looks as if you are not using the myrinet transport
at all and tcp/ip instead. i would check the executable and
consider a recompile with the proper settings (and an update).

XB> yet another related question
 from version 3.11.1, is it possible that
XB> elapsed time is incorporated to CPU time? I say this, because I find both
XB> times suspiciously very similar!!

with a good network this should be the case. also some MPI 
implementations (MPICH is a known offender) use a method called 
'busy-looping' to potentially reduce the latency (at least with
badly written tcp/ip stacks). this will consume a lot of cpu
time, unlike calling sched_yield() and waiting on select().


cheers, 
   axel.

XB> 
XB> thank you!
XB> 
XB>  
XB> 
XB>  
XB> 
XB> Xevi  =o)
XB> 
XB>  
XB> 
XB>  
XB> 
XB> 

-- 
=======================================================================
Axel Kohlmeyer   akohlmey at cmm.chem.upenn.edu   http://www.cmm.upenn.edu
   Center for Molecular Modeling   --   University of Pennsylvania
Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
tel: 1-215-898-1582,  fax: 1-215-573-6233,  office-tel: 1-215-898-5425
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.