Hi John --
Thanks for the feedback (and I agree with your comment about the
multiplicity of library needs 100%).
I should clarify that in asking for users to consider contributing to the
library population effort (whether yourself or others reading this), the
approach we're taking is to wrap/interoperate with existing libraries
rather than port them to Chapel (so far...). That is, it is something
that can fit a user, rather than developer, role. In our experience,
someone familiar enough with Chapel to program in it productively can
write the interoperation declarations and shim code required to wrap a
library in most cases (and we've got experience in this area that we can
share to help.., before long, I'd like to start putting together a "best
practices for wrapping libraries in Chapel" document to capture some of
that knowledge and experience).
As a specific example, the FFTW case I mentioned before was primarily
contributed by a user who was interested in using FFTW from within Chapel.
It ended up being a joint effort as we looked for ways to make the
interfaces more Chapeltastic, but having his experience and real-life use
cases (as opposed to our guesses as to what those may or may not be) made
for a really strong partnership and a nice result (I think).
Best wishes,
-Brad
On Thu, 13 Aug 2015, John Hall wrote:
Thanks for the thorough reply Brad.
In terms of "next libraries", I would imagine that if you ask 3 people
you'd get 4 answers. I imagine that the developers have their hands full.
I'm more of a user than a developer..At best, I can let you know things in
other languages that I commonly use that would be very important for me
transitioning to Chapel. I would imagine others have similar requirements.
For instance, when I use python, numpy, scipy, and pandas are a given (R or
Matlab I use similar facilities). I often use some kind of database (SQL).
I commonly call nlopt from other languages. I would imagine that
interoperability with other languages would also make it more likely for
people to use Chapel. I see a pyChapel package. This seems like a step in
the right direction.
On Thu, Aug 13, 2015 at 8:34 PM, Brad Chamberlain <[email protected]> wrote:
Hi John --
1) The latest benchmark I can find (also on the Chapel website) is from
October 2013. It does not show particularly strong performance. I imagine
that there has been a lot of improvements since then. Are there any more
recent performance evaluations?
Things have improved a lot since 2013. The best sources for recent
performance results are the release notes that we've been putting together
for the past several releases (so that's an every-six-months report):
http://chapel.cray.com/download.html#releaseNotes
Results may appear in various decks for each release, but for the latest
release, most of them appear in the "Performance Optimizationsn and
Generated Code Improvements" deck and the "Benchmarks and Perforamnce
Results" deck. The other place to look for "real-time" data are the
performance graphs that we plot as part of our nightly regression testing:
http://chapel.sourceforge.net/perf/
2) Does/will Chapel support SIMD? What about GPUs?
As of version 1.11, Chapel's forall loops code generate for gcc and icc as
loops with vectorization hints that lead to vectorization. We're still in
the process of evaluating how our performance compares to hand-generated C
loops for various idioms (using the LCALS benchmark suite from LLNL), but
we have seen improvements for some cases. You can read more about this
capability in the same performance deck mentioned above, or in Elliot
Ronaghan's CHIUW 2015 talk:
http://chapel.cray.com/CHIUW2015.html
On the master branch, we've also recently introduced a vectorizeOnly()
iterator which has the following effect:
forall i in 1..n // uses multitasking; tries to vectorize
while:
forall i in vectorizeOnly(1..n) // only tries to vectorize
This will be in the upcoming 1.12 release this fall.
As for GPUs, Albert Sidelnik at UIUC did some early work on Chapel
execution on GPUs that was captured in an IPDPS paper:
Performance Portability with the Chapel Language. Albert Sidelnik,
Saeed Maleki, Bradford L. Chamberlain, María J. Garzarán, David
Padua. IPDPS 2012, May 2012.
but unfortunately this work did not make it back into the master branch.
Enabling GPU support on the master branch is part of future work, but
currently queued up behind good support for Intel Xeon Phi.
3) I see a lot of development, at various stages, on parallel linear
algebra libraries. For instance, PBLAS, CLBLAS, ScaLAPAK, PLASMA, MAGMA. I
did some cursory search for Chapel's support for either these libraries
(like maybe with a wrapper to C or something) or similar functionality
natively, but I didn't see anything. If this support is available, can you
direct me to some of it. Alternately, what is the timeline for Chapel to
have comparable functionality (by whatever means)?
This has become an increasingly common request from users and has become
more of a priority for the team in the past year (support for standard
libraries). One of our interns spent the summer working on creating a
Chapel interface to LAPACK which we expect to be committed to the master
branch this week. An external developer is working on wrapping BLAS and we
hope that will be contributed in time for this fall's release, but it's TBD
(it will show up on master as soon as it is available in any case). You may
have found that we already have wrappers for some key FFTW routines:
http://chapel.cray.com/docs/latest/modules/standard/FFTW.html
We're interested in having the community contribute such wrapped libraries
back to the project given the number of them (FFTW was a community
contribution for example), and are continuing to take requests for "next
libraries" to wrap ourselves as reources permit.
Thanks for your interest in Chapel and please let us know if you have
any further questions,
-Brad
------------------------------------------------------------------------------
_______________________________________________
Chapel-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/chapel-users
