new SGI Origin & Onyx 3k?

[Robert G. Brown]

On Thu, 27 Jul 2000, Eugene Leitl wrote:

> Greg Lindahl writes:
> 
>  > The Beowulf Underground site already has a place that clusters owners can
>  > put info like that, but few people use it. How do you think you could
>  > organize such a poll so that you'd get a lot of responses?
> 
> I think we should first amass a number of questions to ask (either
> onlist or offlist), weed out the less relevant, and after the result
> is commonly considered satisfactory, publicize it on as many mailing
> lists (CCL, etc.) and web sites (top500.org, supercomputer.org, etc)
> as possible.
> 
> Then, we can only hope and wait that the users invest a few minutes of 
> their time.

I second this approach.  What is needed is:

  a) A real database set up in the back end (e.g. mysql).
  b) This requires a (preferrably extensible) database design with all
     the descriptors Eugene refers to collected in e.g. name-keyed records.
  c) A pair of web (XML/SQL/Java?) front ends -- one for entry and one
     for browsing, keyword searches, and to facilitate entry into
  d) summary display tools that construct and show e.g. performance
     graphs and histograms.

For the latter to be useful they have to be based on a uniform standard
of microbenchmarks, and the standard itself needs to be both open source
and "controlled" in the sense that it discourages tuning and
cherrypicking.  I'd strongly recommend lmbench (www.bitmover.com) by
Larry McVoy and Carl Staelin, as I've been working with the two of them
for several months trying to turn this tool into something suitable for
boilerplate use in beowulf design anyway.  This is basically a suite of
microbenchmark tools that measure e.g. memory latencies and bandwidths,
file latencies and bandwidths, network latencies and bandwidths, and
latencies/rates for a variety of systems calls and (just being added
now) elementary e.g. numerical instructions.  It generates a "report"
that is standardized and can be shaped with a bit more work into a
comprehensive database entry for specific node/network combinations that
can be searched, graphed, and sanely compared.  Because the tool has
been around for a while, results can even be compared across time and
architecture/OS combinations.

For what it's worth, Linus and the kernel developers seem to rely on
lmbench microbenchmarks to measure all sorts of things as they work.  It
is a quality toolset and would open an objective performance analysis
channel between beowulf folks, kernel folks, and even compiler folks
(compiler weaknesses or bugs are one of many things it can reveal).

Something like this website was proposed a few months ago and is a truly
wonderful idea, but it requires nontrivial resources to build.  As you
can see, SQL, XML, Java/Perl/Python/Whatever (I'm not religious) tools
and skills, and likely help collecting the benchmark tools (which need
to be linked/distributed from the performance page).  I have most of
these skills but am quadruple parked as it is -- it's all I can do to
continue working on lmbench.

If somebody will undertake the responsibility for the core web app and
DB, I'd be happy to work with them and contribute what I can.  As far as
hosting is concerned, I personally think that this belongs square under
www.beowulf.org.  I don't know if scyld.com has humans who can
contribute and maintain the core web/DB, though, and as long as it is
linked back there it could likely live anywhere.

Note well that this site would benefit everybody.  Beowulf owners,
beowulf designers, turnkey vendors, app designers/vendors, and systems
level developers.  I don't think it would be hard to get folks to enter
data into a form, especially if they come across the form while
obtaining detailed and consistent information on large numbers of
beowulf/cluster systems.

   rgb

> 
> _______________________________________________
> Beowulf mailing list
> Beowulf at beowulf.org
> http://www.beowulf.org/mailman/listinfo/beowulf
> 

Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb at phy.duke.edu