> Hi All.
>
> I just wonder that which of allgatherv and reduce_scatter is faster?
> Basically, they deal with same amout of data but in opposite way.
> In my experience, reduce_scatter is much slower that allgatherv.
> I don't know how they work differently in implementation level.
> Can anybody help me understand the speed difference of allgatherv and
> reduce_scatter?
In LAM, Allgatherv is implemented using N calls to Gatherv while
Reduce_scatter is implemented using a Reduce followed by a Scatterv. You
might want to see the source code (/share/ssi/coll/lam_basic/src) for
above collective operations.
When you say that reduce_scatter is slower, it depends upon how you
measure the completion time, the message size and the number of processes.
I think that for large number of processes and sufficiently large message
sizes, allgatherv becomes slower compared to the reduce_scatter because of
the implementation. Further, in allgatherv, rank0 finishes early compared
to rank1 and so on. In reduce_scatter all the processes finish almost at
the same time.
Hope this helps.
> Thank you so much
>
> Byungchan Kim
>
> _________________________________________________________________
> Fast, faster, fastest: Upgrade to Cable or DSL today!
> https://broadband.msn.com
>
> _______________________________________________
> This list is archived at http://www.lam-mpi.org/MailArchives/lam/
>
|
