[mpich-discuss] Latency or the B/W, which is the bottleneck for most MPI codes?
Si Hammond
simon.hammond at gmail.com
Fri Mar 5 15:20:33 CST 2010
Hi Junli,
I'm afraid that this depends on lots of aspects of the job being run:
- Algorithm or code being used (message decomposition etc)
- Underlying architecture (what is the latency of the network etc?)
- Code configuration (e.g. are problems too small to achieve good parallel efficiency etc).
Something else to consider is the time spent in the MPI stack, this can be quite a significant part of the cost in a message transfer.
In my experience with some large codes, lots of the users will try to target a processor and application configuration so that the factors above are "optimised" for the run.
S.
On 5 Mar 2010, at 20:55, junli gu wrote:
> Hi everyone:
>
> I am wondering to most MPI codes, which is the performance bottleneck, the latency or the B/W?
>
> I am doing research to optimize the cache-coherent shared memory architecture to run MPI codes more efficiently. Although I am trying to reduce the latency for message data transferring, then I realized that maybe the B/W is the real bottleneck. I have heard the comments that most well-written MPI codes are not latency-sensitive but B/W sensitive. So instead of making single message transfer faster, increasing the B/W to transfer the chunk of data faster is the right way to improve performance.
>
> Can you give us some comments about this?
>
> Thank you in advance!
>
> --
> ************************************************
> Junli Gu--谷俊丽
> Coordinate Science Lab
> University of Illinois at Urbana-Champaign
> ************************************************
> _______________________________________________
> mpich-discuss mailing list
> mpich-discuss at mcs.anl.gov
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
---------------------------------------------------------------------------------------
Si Hammond
Research & Knowledge Transfer Associate
Performance Modelling, Analysis and Optimisation Team
High Performance Systems Group
Department of Computer Science
University of Warwick, CV4 7AL, UK
http://go.warwick.ac.uk/hpsg
----------------------------------------------------------------------------------------
More information about the mpich-discuss
mailing list