[mpich-discuss] Determining latency of MPI communication routines

[James Dinan]

Hi Vaibhav,

Accurately benchmarking collective operations is tricky.  Throughput and 
latency are often measured differently, so bandwidth numbers aren't 
necessarily convertible into latency numbers.  This is because 
collective operations issued one after another will overlap with one 
another and effectively pipeline.  There have been several papers on 
this topic that you might find helpful:

http://www.springerlink.com/content/7ygll9u0h02t8mth/
http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.149.3571

Best,
  ~Jim.

On 3/14/11 9:52 AM, vaibhav dutt wrote:
> Hi all,
>
> This question might seem out of place but as I was not getting a reply
> from the LLC bench
> discussion group, I though I might get an answer here.
> I am trying to determine latency of MPI communication routines through
> LLC bench tool.
> For MPI_reduce etc. it gives the message size and KB/s in the graph of
> the results.
> For finding the latency of the operation for a given message size I am
> dividing the message size by
> the transfer rate corresponding to that message. Is this the correct way?
>
> Thank You,
> Vaibhav
>
>
>
> _______________________________________________
> mpich-discuss mailing list
> mpich-discuss at mcs.anl.gov
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss