Performance analysis of CP2K 2.6 vs 2.7

[Abhishek Bagusetty]

Hi Andreas,

You are right there could be an issue with the number of MPI ranks used for 
the test cases. Initially, I had experiments performed with 128 MPI ranks 
for all the 3 tests cases and received a warning suggesting to use squared 
MPI ranks and hence 2.7 (tuned and untuned) test case was performed with 
121 ranks. The conclusion is that there was a very minor difference with 
the use of 128 or 121 for version 2.7 and still they significantly slower 
wen compared to 2.6.

2.6 (fastest) : 128 MPI ranks
2.7 tuned with LIBSMM : 121 ranks
2.7 w/o LIBSMM : 121 ranks

Abhishek

On Monday, June 29, 2015 at 3:43:01 AM UTC-4, Andreas Glöss wrote:
>
> Hi Abhishek,
>
> I just had a quick look at your outputs. You are not using the same number 
> of MPI ranks for your experiment (Total number of message passing 
> processes: 128/121), maybe some default was changed?
> Even thought the larger, non-square number of MPI ranks is not optimal for 
> MPI performance, it might improve local multiplications and pdsyevd due to 
> smaller data portions.
>
> Andreas
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20150629/0e0eae62/attachment.htm>