[mpich-discuss] What are the ways to improve efficiency of node?

wangxinquan wangxinquan at tju.edu.cn
Mon Apr 14 19:52:09 CDT 2008


Dear all,

   Recently, I'm doing a vc-relax calculation of 96 atom system(2*2*2 supercell,24cpu).
   The load for each node is as follow:  
   
   node(2cpu work/node):
   281:98%
   325:84%
   337:73%
   280:87%
   307:77%
   276:64%
   318:55%
   351:64%
   341:50%
   329:23%

   node(1cpu work/node):
   322:93%
   269:100%
   343:76%
   300:99%

   node329 and node341 are inefficient.
   My questions are:
   1. What are the ways to improve efficiency of node?
   2. By now 8 iterations have been taken 10 hours. I'm afraid it is a little slow.
      Is 24cpu enough for 96 atoms relaxation?

PS:
------------------------------------------------------------
HPC background:
Nankai Stars (http://202.113.29.200/introduce.htm)
800 Xeon 3.06 Ghz CPU (400 nodes)   
800 GB Memory    
53T High-Speed Storage    
Myrinet
Parallel jobs are run and debuged through Platform LSF system.
Mpich_gm driver:1.2.6..13a
Test package: Espresso-3.2.3
-------------------------------------------------------------

Any hints will be deeply appreciated!
Best regards, Wang

=====================================
X.Q. Wang
wangxinquan at tju.edu.cn
Schoolof Chemical Engineeringand Technology
TianjinUniversity
92 Weijin Road, Tianjin, P. R. China
tel:86-22-27890268, fax: 86-22-27892301
=====================================




More information about the mpich-discuss mailing list