Re: [mpi-21] Proposal: Non-blocking collective operations for MPI-3

["Anthony Skjellum" <tony@xxxxxxxxxxx>]

Hi. The implementation as layered library is not a primary concern for a standard.  A reason for saying collective is layerable on point to point is useful in re semantics of order, naming of endpoints (communicator context) and for discussing total ordering over a group, but implementations shouldn't care that much when there is acceleration hardware...


Anthony Skjellum, PhD
Professor and Chair
Dept. of Computer and Information Sciences
University of Alabama at Birmingham

-----Original Message-----
From: Dhabaleswar Panda <panda@xxxxxxxxxxxxxxxxxx>

Date: Wed, 16 Jan 2008 18:06:21 
To:mpi-21@xxxxxxxxxxxxx
Cc:Dhabaleswar Panda <panda@xxxxxxxxxxxxxxxxxx>
Subject: Re: [mpi-21] Proposal: Non-blocking collective operations for MPI-3


> > I don't see how non-blocking collectives can be implemented efficiently
> > on top of MPI, indeed it would be equally hard to implement blocking
> > collectives outside of the library.
> Many MPI collectives that I know of (the MPICH collectives as well as
> the tuned Open MPI collectives and most of the collectives in MVAPICH to
> name some well-known implementations) are implemented on top of
> point-to-point messages. So this would also be possible on top of MPI.
> LibNBC implements non-blocking collective operations on top of MPI, the
> implementation and a performance analysis of it has been presented at
> the Supercomputing Conference (SC07) last year [1].

For clarification, several collectives in MVAPICH/MVAPICH2 have optimized
shared-memory-based designs to exploit emerging multi-core architectures.
Thus, they are not implemented on top of point-to-point messages.

Thanks,

DK

> > I'd like to see non-blocking collectives in the standard however without
> > the restriction that MPI_* should match MPI_I*, MPI_Barrier should be a
> > distinct operation from MPI_IBarrier/MPI_Wait.
> Why do you want to handle the barrier operation apart from the others?
>
> Best,
>   Torsten
>
> [1]: "Implementation and Performance Analysis of Non-Blocking Collective
> Operations for MPI" In proceedings of the 2007 International Conference
> on High Performance Computing, Networking, Storage and Analysis, SC07,
> presented in Reno, USA, IEEE Computer Society/ACM, Nov.  2007
>
> --
>  bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ -----
> Indiana University    | http://www.indiana.edu
> Open Systems Lab      | http://osl.iu.edu/
> 150 S. Woodlawn Ave.  | Bloomington, IN, 474045-7104 | USA
> Lindley Hall Room 135 | +01 (812) 855-3608
>