[Mpi-forum] MPI-3 One-Sided Communications

[Anthony Skjellum]

Hi, I am not familiar with you project.  Tell us more.

In order to achieve lowest latency (or overhead, depending on your optimization, or a blend thereof), a protocol designer might not want to pay for fixed costs that once amortized over long transfers yields more long transfer performance.  Classic
Fixed vs variable cost situation.

The classic trades in two-sided are zero copy for long and two copy for short.

If you have evidence to the contrary, great... Lots of people reported in the past the effect of a tradeoff as I described coming from the complex semantics of the mpi one sided api.

One also has to ask if we can get even lower latency with less complex protocol too imho.

Do you get both lower latency than two sided short msgs and higher bandwidth than two sided?
What is the baseline of performance for two sided? Is it optimized? :-)


Tony

Anthony Skjellum, PhD

Verari Systems, Inc
+1-205-807-4968



-----Original Message-----
From: Jeff Hammond <jeff.science at gmail.com>

Date: Thu, 23 Apr 2009 20:03:07 
To: Main MPI Forum mailing list<mpi-forum at lists.mpi-forum.org>
Subject: Re: [Mpi-forum] MPI-3 One-Sided Communications


What does "allow for potentially lower bandwidth on long transfers as
a side effect" mean?  That one-sided should have lower latency at the
expense of lower bandwidth?  That's an unnecessary and intolerable
compromise.  ARMCI proves that one-sided bandwidth can equal if not
beat MPI (http://www.emsl.pnl.gov/docs/parsoft/armci/performance.htm).

Jeff

On Thu, Apr 23, 2009 at 7:32 PM, Anthony Skjellum <tony at cis.uab.edu> wrote:
> Hi, I was giving my opinion...
>
> The main requirements for one sided was achieve lower latency than two sided (as well as one sidedness), and allow for potentially lower bandwidth on long transfers as a side effect... What one sided achieved is higher latency and higher bandwidth in typical (quality) implementations.  The api is consequently a mismatch to its original purpose... It diverged from its design center.. Missed the mark... Way too complex compared to simple put and get.
>
> Starting again from first principles with the goal of very low latency remote put and get without the legacy of existing one sided is my suggestion.
>
> In other words : Start over.
>
> Perhaps that is why now we have an active messages group?
> Not sure.
>
>
>
> Tony
> Anthony Skjellum, PhD
> Professor and Chair
> Dept. of Computer and Information Sciences
> University of Alabama at Birmingham
> +1-205-807-4968
>
> -----Original Message-----
> From: Vinod tipparaju <tipparajuv at hotmail.com>
>
> Date: Thu, 23 Apr 2009 20:09:38
> To: <mpi-forum at lists.mpi-forum.org>
> Subject: Re: [Mpi-forum] MPI-3 One-Sided Communications
>
>
>_______________________________________________
> mpi-forum mailing list
> mpi-forum at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum
>
>
>_______________________________________________
> mpi-forum mailing list
> mpi-forum at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum
>



-- 
Jeff Hammond
The University of Chicago
http://home.uchicago.edu/~jhammond/

_______________________________________________
mpi-forum mailing list
mpi-forum at lists.mpi-forum.org
http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum