I agree with the general tone of caution in Bill's note but I don't
think waiting until just before SC96 will significantly increase my
confidence in the proposal. That is just not enough time.
I'm mostly concerned with distributed memory so I don't have the kind
of specific objections that Eric and others have raised. I have few
technical objections to anything in MPI-2 vis-a-vis LAM. We have
a powerful daemon and a powerful underlying OS - we can implement
most anything, though I do not relish the mandated work. I would rather
be doing something else to build MPI than implementing MPI-2.
Beyond dynamic processes, a few other misc. issues and maybe some
external interfaces, we will definitely have a "wait and see" attitude
toward MPI-2.
I am interested in the evolution of parallel computing technology
and MPI's impact on it.
MPI-1 was very much standardizing current practice. It was overdue.
The only obvious marketing weakness of MPI-1 was dynamic processes.
Lots of people (especially PVM start-up coders) who think they need it
don't really need it. Moreover, I have started thinking that people
who want to program heterogeneous distributed systems (like factory
stuff or toaster/lights/doorbell stuff) might be
better off with PVM. Still, marketing rules and MPI must have dynamic
processes.
The rest of MPI-2 seems to me, frankly, like an academic research
project. This is fine for a .edu, but not for a Forum with serious
impact potential on the industry at large. Look at the similarities.
All the current proposals are high quality and represent a great effort,
albeit by a limited number of people [how many people who will vote
1-sided into the standard have put their all into shaping and validating it?].
The work is far more original than standard practice. There is probably
enough for a Phd for the principle authors, if they need an extra one.
The motivation is more "here's something that should be done" rather than
the aforementioned motivation for dynamic processes. The justification
is more "lots of users I know" rather than the PVM/P4/PARMACS/Express
situation we had with message-passing.
1-sided and MPI-IO demonstrate that MPIF is a forum for standardizing
anything relevant to parallel computing, not just message-passing.
We have become standard-happy.
MPI-1 is barely a toddler. It needs to be nurtured with tools and
implementations. Companies like SGI and HP are just now getting their
implementations on-line while at the same time the Forum is raising
the bar.
What the industry needs, IMO, is lots of MPI-IO-like projects before
they got the gleam of standard in their eyes (maybe they had it at
the beginning). I mean multiple projects in each of the areas, just
like message-passing was. Then we can sit down in a few years and
repeat the MPI-1 exercise. We have definitely got the cart before the
horse in MPI-2 today.
There is every good chance that serious glitches will pop up in
1-sided after a larger number of implementations are attempted.
It has nothing to do with the care and quality that are obvious in
the current proposal. It is just a risky low-level topic.
Is portability enough of a reason to target my shared memory
application to MPI-2 instead of native facilities on my expensive new
Power Challenge? I only see 1-sided as a win on something like
a T3D where the model fits so well and portability is icing on the cake.
I would not assume that the Forum has the last say on MPI-2.
To avoid the standards-that-never-made-it file, there will have to
be implementations for all the major w/s vendors, perhaps through
a PD implementations, and for the current supercomputer elite.
If enough people "wait-and-see", and the substantial offering of
IBM does not begin a panic attack, the whole thing could stall.
So I urge everyone to vote against most everything except the dynamic
chapter. This doesn't mean the proposals are trashed. They just
become research projects in name as well as substance.
-- Greg