Analysis of performance of parallel code as processors increase

[Ben Tay]

Hi,

I have coded in parallel using PETSc and Hypre. I found that going from 
1 to 4 processors gives an almost 4 times increase. However from 4 to 8 
processors only increase performance by 1.2-1.5 instead of 2.

Is the slowdown due to the size of the matrix being not large enough? 
Currently I am using 600x2160 to do the benchmark. Even when increase 
the matrix size to 900x3240  or 1200x2160, the performance increase is 
also not much. Is it possible to use -log_summary find out the error? I 
have attached the log file comparison for the 4 and 8 processors, I 
found that some event like VecScatterEnd, VecNorm and MatAssemblyBegin 
have much higher ratios. Does it indicate something? Another strange 
thing is that MatAssemblyBegin for the 4 pros has a much higher ratio 
than the 8pros. I thought there should be less communications for the 4 
pros case, and so the ratio should be lower. Does it mean there's some 
communication problem at that time?

Thank you very much.

Regards


-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: test4_600_29min.txt
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20080607/08482424/attachment.txt>
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: test8_600_19min.txt
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20080607/08482424/attachment-0001.txt>