Discussion:
Critique of OpenMP ......
(too old to reply)
Solomon_Man
2007-11-10 18:03:26 UTC
Permalink
All,
My partner and I are working on a research paper for a graduate course
and wanted to see what short comings that OpenMP has compared to MPI.
As a student we are especially interested in its short comings in a
commercial environment as we may not be exposed to this in an academic
atmosphere.

We are also interested in portability issues, as the academic studies
that we have ran across have shown conflicting results. We would be
interested in anything anyone would have to say about OpenMP.

Thanks,
Saranga and Chris

P.S. Does anyone know how to determine OpenMP version on a Solaris
machine?

--
Justin W
2007-11-11 20:58:10 UTC
Permalink
[C.p. moderator: again, may be posting delays while I am in SC'07.]
Post by Solomon_Man
and wanted to see what short comings that OpenMP has compared to MPI.
From the standpoint of viability in the software industry... I
generally think the technology with the most responsive support wins.
This support could come from an outside vendor, open source community
or even internal employees at the company. I work for a company that
builds a lot of stuff internally because then we see the benefit of
taking ownership of the performance tweaking, bug-fixing etc. The
open source community is good a way to get "free" support, but you
won't always get the responsiveness or quality (not to start an
argument) that your business may rely on. In terms of vendor
support... they can be wishy washy as well. My company has worked
with some all-star vendors out there... and we've worked with some
duds. I cannot comment specifically to the quality of vendor support
for OpenMP or any implementation of MPI. My personal favorite is
building these components internally. You get the most responsive
support but at the highest cost (employee salary, time lost to ramping
up new hires on existing technology etc). You also have execution
risk... ie, what if I can't produce something as reliably as a vendor
can. But in the end, you own that software and the architecture it
runs on.

So... now to answer your question. I personally... just from a
support standpoint would tend to prefer MPI. After all, MPI is just an
interface, the guts can be built internally to do whatever. Heck,
even tweak the performance of the guts to work specifically on your
architecture and throw away the overhead you'd never use anyway.


--
Solomon_Man
2007-11-13 21:55:37 UTC
Permalink
Post by Justin W
Post by Solomon_Man
and wanted to see what short comings that OpenMP has compared to MPI.
From the standpoint of viability in the software industry... I
generally think the technology with the most responsive support wins.
This support could come from an outside vendor, open source community
or even internal employees at the company. I work for a company that
builds a lot of stuff internally because then we see the benefit of
taking ownership of the performance tweaking, bug-fixing etc. The
open source community is good a way to get "free" support, but you
won't always get the responsiveness or quality (not to start an
argument) that your business may rely on. In terms of vendor
support... they can be wishy washy as well. My company has worked
with some all-star vendors out there... and we've worked with some
duds. I cannot comment specifically to the quality of vendor support
for OpenMP or any implementation of MPI. My personal favorite is
building these components internally. You get the most responsive
support but at the highest cost (employee salary, time lost to ramping
up new hires on existing technology etc). You also have execution
risk... ie, what if I can't produce something as reliably as a vendor
can. But in the end, you own that software and the architecture it
runs on.
So... now to answer your question. I personally... just from a
support standpoint would tend to prefer MPI. After all, MPI is just an
interface, the guts can be built internally to do whatever. Heck,
even tweak the performance of the guts to work specifically on your
architecture and throw away the overhead you'd never use anyway.
Justin,
Thanks for your response.
I believe one of openmp's biggest strengths is its adoption by the
programming community. My guess this adoption is based on ease of use.
I have used both MPI and openMP in academia but never in a commercial
atmosphere. I have been in the commercial arena for the past 10 yrs,
but never have run across the reason to use either.

Thanks again for your response,
Chris


--
David Golden
2007-11-14 23:21:12 UTC
Permalink
We would be interested in anything anyone would have to
say about OpenMP.
Well, do note that it's not an either/or thing. Don't forget it's quite
possible to use both OpenMP (at least non-"Cluster OpenMP") and MPI in
the one program... if you want to. Some people look into it
for "cluster of SMPs" situations. Whether it's worth it is another
matter, but there are papers out there a search engine query away.

--

Loading...