[Mpi3-rma] Updated MPI-3 RMA proposal 1
balaji at mcs.anl.gov
Wed Jun 16 21:48:27 CDT 2010
Btw, there was also some discussion on whether we can use MPI_Win_fence
to achieve the same thing, but the problem was that we couldn't use
flush/flushall in the Win_fence epoch; that was only relevant for the
So, basically, we wanted something like Win_fence, but could be used
within the lock/unlock epoch.
On 06/16/2010 09:37 PM, Pavan Balaji wrote:
> We had some discussion on the point that allflushall was not one-sided
> like the rest of the passive target calls. But the consensus was that we
> were really looking for functionality -- if it doesn't match the rest of
> passive target, so be it. We can stop calling it "passive target" and
> rename it to something else, if it helps.
> -- Pavan
> On 06/16/2010 09:09 PM, Underwood, Keith D wrote:
>>> This ship has sailed. You're the only person who has voiced an objection to all_flush_all and numerous others
>>> have demonstrated at length why it is necessary. I do not find the obsession with pure-passive target
>>> semantics to be a compelling reason to reject all_flush_all.
>> All_flush_all will further dissuade good implementations of passive-target semantics. To my understanding, many implementations don't provide compliant passive-target semantics now and the introduction of all_flush_all will give another out to those implementations. So, yeah, call me a purist, but it isn't JUST about doing API design that looks reasonable and has reasonable names.
>>>> Do we have the use cases for both of those and some attempt at
>>>> quantification of the performance advantage in some implementation?
>>> Yes. I will not repeat the use case for all_flush_all since we have
>>> discussed that topic far too long already. Somewhat else should
>>> present the multiple locks issue generically before I give the Global
>>> Arrays use case.
>> I have REPEATEDLY asked for a quantification of the advantage, and NOBODY has offered a quantitative discussion. Lots of hand waving and discussion of challenges that are faced on BG, even an interesting implementation case given by Brian Smith, but he agreed that there would be a crossover where all_flush_all could be slower than flush_all + barrier, since there is an iteration that is involved. And, right now, there is only one system I know of where there is an opportunity for an advantage: BG. So, many of us CAN'T quantify the advantage.
>> mpi3-rma mailing list
>> mpi3-rma at lists.mpi-forum.org
More information about the mpiwg-rma