Re: feedback on a new MPI idea
(no name) (j_nieplocha@pnl.gov)
Wed, 28 Feb 1996 14:25:59 -0800
In message <199602282200.OAA10880@win233.nas.nasa.gov>Bill Saphir writes:
> ........... Buffering is evil
> because the i860 has such poor memory bandwidth. (memory-to-memory
> copies used to go at 4.8 MB/s, and sometimes still do!). IMO MPI
> ready send was a mistake because its main use appears to be to get
> around deficiencies in the underlying MPI implementation (as
> above). An MPI implementation that relies on NX can't solve the
> buffering problem.
>
I am not arguing with your point about the NX protocols; however, your estimate
of the i860 memory bandwidth is a bit too pessimistic. People get around 90MB/s
when copying alligned data on Paragons.
Jarek Nieplocha