[mpich2-dev] Fwd: MPICH2 - 2012 SPXXL requirements

[Jeff Hammond]

I have proposed the following for inclusion in the SPXXL _requirements_ for IBM.

If you have comments, I need them by the end of the week.

Best,

Jeff


---------- Forwarded message ----------
From: Ray Loy <rloy at alcf.anl.gov>
Date: Wed, May 30, 2012 at 10:25 AM
Subject: Re: MPICH2 - 2012 SPXXL requirements
To: Jeff Hammond <jhammond at alcf.anl.gov>
Cc: Susan Coghlan <smc at alcf.anl.gov>, Kalyan Kumaran
<kumaran at alcf.anl.gov>, ALCF Apps <alcf-apps at alcf.anl.gov>

This sounds ok to me.  Forwarding to apps for comments.

Ray

----- Original Message -----
> From: "Jeff Hammond" <jhammond at alcf.anl.gov>
> To: "Ray Loy" <rloy at alcf.anl.gov>
> Cc: "Susan Coghlan" <smc at alcf.anl.gov>, "Kalyan Kumaran" <kumaran at alcf.anl.gov>
> Sent: Friday, May 25, 2012 10:08:26 PM
> Subject: Re: MPICH2 - 2012 SPXXL requirements
> Title: Tighter MPICH2 integration on BG and MPI-3 support
>
> The Blue Gene/P implementation of MPI was branched off of what is now
> an obsolete version of MPICH2 (1.1), which means that BG/P lacks numerous
> features that have long since become ubiquitous, especially MPI 2.2 support,
> which includes C99 datatypes, MPI_Reduce_scatter_block, etc.
> Furthermore, more recent versions of MPICH2 have optimizations and bug fixes
> that would resolve bugs in BGP-MPI that are prohibitive to fix without moving
> it to these later versions. When the MPI-3 standard is released in
> approximately six months, BGQ-MPI will already be obsolete since it supports
> only MPI 2.2 as implemented in MPICH2 1.4.1p1.
>
> IBM should pursue a development model for BG-MPI that makes it easier to
> keep up-to-date with the latest release of MPICH2, which will almost certainly be
> the first implementation compliant with new versions of the MPI standard
> (this was true for many, if not all of the past versions of the MPI standard).
> This could be accomplished either by restructuring of the code to more
> tightly integrate with MPICH2 below the device layer; or alternatively
> by a sustained development effort to merge BG-specific code every MPICH2 release.
> SPXXL notes that the former would be a better investment.
>


-- 
Jeff Hammond
Argonne Leadership Computing Facility
University of Chicago Computation Institute
jhammond at alcf.anl.gov / (630) 252-5381
http://www.linkedin.com/in/jeffhammond
https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond