Yes, in case of limited memory the preferred mode of subset operation would be to allow implementations reflect what part of MPI-3 they do support, rather than let application pick and choose.

However, in a different context the concept of the subsets may be open to other uses, for example, by those who can shoulder the validation cost and justify this by increased performance of the subset combinations they offer.

From: mpi-22-bounces@lists.mpi-forum.org [mailto:mpi-22-bounces@lists.mpi-forum.org] On Behalf Of Richard Treumann
Sent: Friday, May 09, 2008 3:12 AM
To: MPI 2.2
Subject: Re: [Mpi-22] Memory footprint concern

Hi Alexander

If we would envision defining subsets in such a way that a machine with limited memory per node and un-fancy OS can clearly document what its MPI offers, maybe it is not a huge addition to complexity. For example if an MPI implementation could say it offers subsets:
MPI 1.2
MPI-IO
MPI-Intercomm Collectives
and nothing else and the meaning would be well defined that may not be too bad.

If you are expecting such a machine to have all or most of MPI-3 available and each user would select only the parts he wants at job launch time it seems to me the complexity for implementation and testing will be prohibitive. In particular if you hope subsets will give better performance because of what they leave out and really be compact because they leave out most unneeded bytes of code the number of permutations seem to me to make that a vain hope.

I have been ambivalent about whether my assertions proposal really has much connection to the subsets concept mostly because I think to be useful assertions must be fairly simple. The only assertion I am 100% sure will pay off and be used is MPI_NO_EAGER_THROTTLE. There are MPI implementations today that act as if the assertion MPI_NO_EAGER_THROTTLE were present on all applications but that is not an option for a vendor MPI because if somebody contacts service and says "Your MPI violates the standard" a vendor like IBM must "fix" it. IBM MPI supports MPI_CANCEL on MPI_ISEND but at least one MPI does not because it is "too expensive". In effect they act like every application has the assertion MPI_NO_SEND_CANCELS. IBM MPI could make use of MPI_NO_REQUEST_MIX if it was available.

Dick

Dick Treumann - MPI Team/TCEM
IBM Systems & Technology Group
Dept 0lva / MS P963 -- 2455 South Road -- Poughkeepsie, NY 12601
Tele (845) 433-7846 Fax (845) 433-8363

"Supalov, Alexander" <alexander.supalov@intel.com>

"Supalov, Alexander" <alexander.supalov@intel.com>

Sent by: mpi-22-bounces@lists.mpi-forum.org

05/08/2008 05:52 PM

Please respond to
"MPI 2.2" <mpi-22@lists.mpi-forum.org>

To	"MPI 2.2" <mpi-22@lists.mpi-forum.org>
cc
Subject	Re: [Mpi-22] Memory footprint concern

Dear Dick,

By the looks of it, MPI-3 is going to be big. Petascale machines may not have OS we're accustomed to, dynamic libraries, and some other things. Smaller system libraries - and smaller MPI - may be needed there. Some of the envisioned MPI-3 features will be needed for some applications, some won't. Same with MPI-2 and MPI-1. Defining subsets may help to open a way to custom cut MPI libraries suitable for particular application classes. How subsets will be implemented is a different matter.

Best regards.

Alexander

Can somebody help me understand this "smaller memory footprint" issue that is part f the subsetting goal better. What systems does it affect? What does "memory footprint" really mean? In the 64 bit address space, virtual address range is not a problem.

On systems I am most familiar with (AIX and I have been told Linux too), if you have a library that contains 1000 subroutines and you run a program than only calls 6 then only the pages that are touched by code for those 6 functions must get placed in real memory. The rest of the object code stays on disk. Program and library text is demand paged. The loading is on page boundries, not subroutine boundries.

With a shared library, if I run a program on a node and touch 6 subroutines and you run a different program that touches those 6 and 10 more then code for all 16 subroutines may be kept in memory but the rest of the library will stay on disk. You and I will share the object code for the 6 subroutines we are both calling.

Someone who wanted to make a libmpi that has MPI-1sided or MPI-IO well isolated in the library structure so simple MPI jobs would not force this extra code into memory could do that today. The user does not need to promise not to call MPI-IO subroutines for them not to to take real memory. The "subsets" would need to be devised by the MPI implementor but would be transparent to the MPI user and not dictated by the standard. The "subsets" the user did not call would remain paged out.

Perhaps all static data defined by the library will come into real memory for each process but is there much reduction from being able to somehow not bring in the static data MPI-IO would require because somebody had promised not to use it?

Dick

Dick Treumann - MPI Team/TCEM
IBM Systems & Technology Group
Dept 0lva / MS P963 -- 2455 South Road -- Poughkeepsie, NY 12601
Tele (845) 433-7846 Fax (845) 433-8363

--------------------------------------------------------------------- Intel GmbH Dornacher Strasse 1 85622 Feldkirchen/Muenchen Germany Sitz der Gesellschaft: Feldkirchen bei Muenchen Geschaeftsfuehrer: Douglas Lusk, Peter Gleissner, Hannes Schwaderer Registergericht: Muenchen HRB 47456 Ust.-IdNr. VAT Registration No.: DE129385895 Citibank Frankfurt (BLZ 502 109 00) 600119052 This e-mail and any attachments may contain confidential material for the sole use of the intended recipient(s). Any review or distribution by others is strictly prohibited. If you are not the intended recipient, please contact the sender and delete all copies._______________________________________________ mpi-22 mailing list mpi-22@lists.mpi-forum.orghttp://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-22