[ofiwg] FW: FW: SC15 Tutorial submission: substantiation of participation

Jeff Hammond jeff.science at gmail.com
Thu May 21 12:49:51 PDT 2015


Regarding Robert's comment, that argument would not work for me.  >90%
of the Top 500 machines support TCP/IP, but that does not justify a
tutorial on that topic.  The vast majority of HPC users do not care
about InfiniBand except for the performance it provides for MPI, etc.

What might have more traction is to bring up the numerous efforts to
find alternatives to MPI, most of which will require a flexible,
high-performance communication conduit.  Currently, many projects
(OCR, Chapel, Legion, UPC++, ...) target GASNet, but OFI may be a more
appropriate target.  (Of course, I personally argue that people should
target MPI-3, but there are good reasons why something closer to the
metal should be used instead.)

>> I didn't read the proposal as I am on the committee, but here's some advice based on previously accepted and rejected tutorials.

I am similarly on the committee and conflicted w.r.t. this submission.
However, I'd review this tutorial favorably if I could :-)

>> I don't think discussing what you say below will help, because those are not likely the people who will attend the tutorial.  If you haven't done so already, explain who would be interested in attending such a tutorial, e.g., MPI developers, PGAS language developers, etc. for all such communities that will be in attendance at SC.  If possible, I'd try make a case for industry participants--they often get forgotten, but they are also the ones who may foot the bill for such a tutorial.  We've had quite a few at the Chapel tutorials in the past.

Not to mention storage and oil & gas folks that today write directly
to IB verbs and are likely interested in OFI for the same reasons.

>> I'd also cite that this is the interface that will be available on CORAL and other such Intel-based systems.  There is a large community of people who will be using this interface.  If Cisco can add some stats about their potential user community, that would be good too.
>> If there's a way to get an international angle on it (e.g., saying something about Intel/Cisco systems in Europe or Asia), that would be good too.

Sorry for being pedantic, but CORAL refers to the DOE procurement
effort to fund 3 systems, 1 of which is being produced by Intel and
Cray.  The Argonne supercomputer referenced here is named Aurora.

>> Finally, if you haven't done so already, cite last year's PGAS tutorial with hard facts.  Attendance numbers (even as a percentage of attendees), slides (though I have to say they were a little lean), length in hours, and feedback (if collected).

The PGAS OFI tutorial was well-received.  I don't have data even
though I was the organizer.  Sameer Shende should have registration
data, which is an upper bound on the number of attendees.

Best,

Jeff


-- 
Jeff Hammond
jeff.science at gmail.com
http://jeffhammond.github.io/



More information about the ofiwg mailing list