If we provide them with a standard ABI, they might reconsider, but
again, as this is not going to be the same for all platforms (Intel's,
IBM's, etc.), they will most likely stay with their kludges for
quite a
while.
Hence, to be serious, the ABI effort should target all popular
platforms.
Best regards.
Alexander
--
Dr Alexander Supalov
Intel GmbH
Hermuelheimer Strasse 8a
50321 Bruehl, Germany
Phone: +49 2232 209034
Mobile: +49 173 511 8735
Fax: +49 2232 209029
-----Original Message-----
From: owner-mpi-21@xxxxxxxxxxxxx [mailto:owner-mpi-21@xxxxxxxxxxxxx]
On
Behalf Of William Yu
Sent: Thursday, January 10, 2008 11:51 PM
To: mpi-21@xxxxxxxxxxxxx
Subject: RE: [mpi-21] ABI - call for working group
I personally have never tried it out on a large suite of applications
yet. But, i presume that these morph layer implementors have some
issues
that could be address by a standards clarification?
On another note, it is probably not standard because it is a "dirty"
workaround?
________________ Reply Header ________________
Subject: RE: [mpi-21] ABI - call for working group
Author: Greg Bronevetsky <bronevetsky1@xxxxxxxx>
Date: January 10th 2008 10:30 pm
A morph layer would need to have code for every MPI oddity out there,
which obviously can get messy. However, all this work only needs to
be done at most once per MPI implementation and probably much less
than that since a lot of MPIs are similar to each other. Furthermore,
if the morph layer was implemented using a modular structure it would
be possible for MPI implementors to submit modules that describe the
oddities of each version of their MPI without being bound into
a particular set of ABI decisions.
The cost of this is somewhat questionable. The cost of creating the
layer is 40-60ns per MPI call on 2.4Ghz Opterons (shown for PnMPI in
their SC07 paper:
http://sc07.supercomputing.org/schedule/pdf/pap224.pdf). Most of the
work required by the morph layer is very basic: rename constants,
copy MPI opaque objects from handles that are used by the application
to the real objects created by MPI, etc. I don't know if PnMPI is
missing support for some oddity of some specific MPI implementation
that will make things a lot more expensive.
As to why they're not a standard yet, its just that these layers are
so new. PnMPI is close to being released and while I'm not sure how
ready MorphMPI is, their sourceforge site claims that its in Alpha
development status. I'd say that the best course of action would be
to identify what parts of the ABI are solved by morph layers, and
what parts are not. A few users who are having ABI-related issues can
evaluate the morph layers to see if they have any fatal flaws. If
they do, then we can standardize the entire ABI proposal. If not, we
can just use the layers and standardize only the parts that we don't
have tools for.
Greg Bronevetsky
Post-Doctoral Researcher
1028 Building 451
Lawrence Livermore National Lab
(925) 424-5756
bronevetsky1@xxxxxxxx
How do morph layers find and load the right MPI library? How do they
manage variously sized opaque (and less opaque) objects? Are there
any
performance penalties when those morph layers are used? Why did not
they
become a de fact standard yet?
Best regards.
Alexander
--
Dr Alexander Supalov
Intel GmbH
Hermuelheimer Strasse 8a
50321 Bruehl, Germany
Phone: +49 2232 209034
Mobile: +49 173 511 8735
Fax: +49 2232 209029
-----Original Message-----
From: owner-mpi-21@xxxxxxxxxxxxx [mailto:owner-mpi-21@mpi-
forum.org] On
Behalf Of Greg Bronevetsky
Sent: Thursday, January 10, 2008 9:38 PM
To: mpi-21@xxxxxxxxxxxxx; mpi-21@xxxxxxxxxxxxx
Cc: mpi-21@xxxxxxxxxxxxx
Subject: Re: [mpi-21] ABI - call for working group
As far as I understand it, a morph layer such as MorphMPI or PnMPI is
a valid implementation of MPI from the point of any application. The
layer itself then chooses the local MPI implementation that it will
use to implement MPI functionality. As such, the application can come
pre-linked to its morph layer and the morph layer can do all the
platform-specific work without the need to recompile anything. While
I still see a need to standardize things like mpirun arguments, which
is something the user actually sees, most of the other use-cases for
the MPI ABI can be satisfied by open-source software that already
exists: MorphMPI and PnMPI.
It seems to me that a more feasible MPI ABI proposal would exclude
any features that solve problems that are already solved by the above
morph layers. Instead, it should focus on things that actually cannot
be done unless they're added to the standard. mpirun arguments is
one. What are the others?
Greg Bronevetsky
Post-Doctoral Researcher
1028 Building 451
Lawrence Livermore National Lab
(925) 424-5756
bronevetsky1@xxxxxxxx
At 10:35 AM 1/10/2008, William Yu wrote:
Quite a detailed justification.
I just curious why writing a "morph" layer is difficult because of
the other libraries that use mpi is in use too?
Does this mean because mathlab needs (example) parallel blas and
both of them need mpi is causing the difficulty?
I see linux device driver folk have the same issue but they seem to
workaround it sufficiently well. Particularly those that involve
both X and the linux kernel. (i worked on a capture card vendor in
the past that distributed binary driver cores and open source
wrappers).
Just curious because fixing an ABI obviously leaves less room for
implementations.
Thanks.
________________ Reply Header ________________
Subject: [mpi-21] ABI - call for working group
Author: Edric Ellis <Edric.Ellis@xxxxxxxxxxxxxxx>
Date: January 10th 2008 5:21 pm
Hi All,
This is a call for organisations to support a working group working
towards defining an ABI for MPI. To make this happen, I need the
support
of 3 other organisations.
I realise that this is a large issue with many facets, many of which
I
don't understand fully (although I have tried to work through the
various sources of information out there - the various mailing list
discussions, and stuff like Greg Lindahl's paper). However, I hope
that
a working group would be able to develop a reasonable proposal.
Below
I
include an outline from the perspective of The MathWorks as to what
we
would like to see:
1. Motivation
Many current users of MPI are perfectly happy to recompile their
application to use different MPIs. However, a growing number of
users
are unable or unwilling to recompile an application to use a
different
MPI. Additionally, parallel processing is becoming increasingly
important to software vendors since they can no longer rely on
increasing single-core performance.
Since it's the case I'm most familiar with, I will describe how
MATLAB
attempts to handle this sort of situation:
1.1 An example application: MATLAB
MATLAB is introducing explicit parallelism as our customers demand
ever
larger data sets and ever higher performance. One part of the value
of
our software is that we bring together various 3rd party libraries
and
put them into a convenient environment for scientists and engineers.
Wherever possible, we prefer to allow our users to substitute their
own
favourite versions of BLAS, LAPACK etc.
1.1.1 Why MATLAB can't currently use all MPIs
Here are the main difficulties encountered when trying to make
MATLAB
work with the widest range of available MPI implementations:
* We do not support customer recompilation of any part of
MATLAB,
therefore any library that we use must be binary compatible with the
one
that we ship
* We do support a "de facto" ABI defined by the implementation
that we ship, several other MPIs match that one (but many don't).
* Adding a "morph" layer is not simple, since we also build on
many other libraries which use MPI (such as BLACS)
* Our choice of "de facto" ABI is limited by the fact that we
only
want to build and qualify one MPI across our 6 supported (commodity)
platforms.
* Aside: we provide means of dealing with the vagaries of
differing mpirun/mpiexec schemes, and have a means of bypassing that
altogether using MPI_Comm_connect/accept.
1.1.2 What MATLAB expects to handle
We fully expect to handle many of the issues relating to switching
libraries:
* we give a flexible means of specifying which binary to
select
* we attempt to avoid making assumptions about the details of
the
MPI implementation (i.e. we stick closely to the standard)
2. Elements of a solution
2.1 Scope
Limit the scope to only those situations where it sensibly applies.
This
may mean that it is not feasible to produce a solution for
applications
written in Fortran due to the following compiler issues:
1. name mangling issues (MPI_INIT vs. mpi_init_ vs. ...)
2. value of .TRUE.
3. issues to do with calling convention
2.2 mpi.h
Define the contents of mpi.h more closely
1. Define values of constants (such as MPI_COMM_WORLD etc.)
2. Define size of MPI_Status
3. Define size and types of MPI handles (such as MPI_Datatype
etc.)
4. Define calling convention (e.g. cdecl vs. stdcall on WIN32
platforms)
2.3 Query implementation in use
It may prove useful to provide a means of querying the MPI layer for
a
description of itself. For example, something along the lines of:
int MPI_Version_info( char * buf, int bufSize )
would allow an application to query at runtime the MPI
implementation
that it is using. The intention is that the output is for
information
only.
3. Pros, Cons and Others
3.1 Pros
1. Allows application developers to ship binaries which can
work
with any MPI implementation
2. MPI developers can test their implementation against those
applications
3. MPI implementors can benefit because of the potential
expanded
user base
4. Implementations already have to support a translation layer for
the
Fortran interface, so there may not necessarily be a large
implementation overhead (ref:
http://www.open-mpi.org/community/lists/users/2005/03/0040.php)
5. Hardware vendors can ensure that their hardware can be used
with
the widest range of applications simply by ensuring that an ABI
conforming MPI is available.
3.2 Cons
1. Standardizing types of MPI handles involves a significant
amount
of work for MPI implementors (but see 3.1.4)
2. Standardizing types of MPI handles may restrict MPI
implementation choices.
3. May not be able to resolve C++ / F90 name mangling issues
3.3 Others
1. Applications still must qualify against each MPI
implementation
that they wish to support.
2. Naming of the shared library - applications may be
reasonably
expected to handle finding the right library using dlopen() or
similar.
Best regards,
Edric.
--
Edric M. Ellis
The MathWorks,
Matrix House,
Cambridge Business Park,
Cambridge CB4 0HH, UK
Tel: +44 (0)1223 226751
---------------------------------------------------------------------
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.
---------------------------------------------------------------------
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.