[mpich-discuss] MPI broadcast question

[William Gropp]

It depends on how the broadcast is being used.  In many cases, minimum completion time is the goal, and the hierarchical communication is preferred.  There are cases where minimum process time is the goal *and* there is flexibility as to timing; in that case, simple linear algorithms can be better.  But for the common case, the hierarchical methods are better because they have relatively low total time and time per process; if the linear algorithm doesn’t satisfy all the rules, it can be very slow.  Note also that for broadcasts that involve more than a small amount of data, the algorithm is usually not a tree but a hierarchical scatter/allgather approach.

Bill

On Mar 15, 2015, at 4:16 AM, YANG Fan <iddmbr at gmail.com> wrote:

> Hi,
> 
> I learnt that MPI does broadcast with some tree-based hierarchical communication, which result in O(log p) latency, where p is the number of processes. However, I'm thinking that if I do the broadcast naively(send to all receivers simultaneously), seems that conceptually the latency can be reduced to O(1).
> 
> Does the above make sense? Thanks!
> 
> Best Regards,
> Fan
> _______________________________________________
> discuss mailing list     discuss at mpich.org
> To manage subscription options or unsubscribe:
> https://lists.mpich.org/mailman/listinfo/discuss

_______________________________________________
discuss mailing list     discuss at mpich.org
To manage subscription options or unsubscribe:
https://lists.mpich.org/mailman/listinfo/discuss