Beginning with beowulf

Robert G. Brown rgb at phy.duke.edu
Thu Feb 27 06:14:28 EST 2003


On Thu, 27 Feb 2003, Santiago Muelas wrote:

> Hello,
> I am a newby to this list and to the beowulf world.

> As it seems to me that the level of this list is quite high, perhaps I
> should find another for "beginners". If so, any suggestion on the
> possible list will be very welcome.

Naaa, this is the right one, and your questions are pretty focused.
Don't worry about it.

> My question is a simple (and I guess a sempiternal) one. I have just
> run my first parallel program in a "beowulf". I have used up to five
> processor in an intranet in this School of Enginnering. Everything run
> o.k. with just one "small" unconvenient: the best result was obtained
> using three processors, but almost the same as with two. Using the five
> was a total disaster.
>
> The reason seems clear. One computer is a dual one. The network is a
> totally standard ethernet 100M.

Don't be TOO hasty to make conclusions.  You could be right, but you
first need to fully understand why this is normal and expected behavior
for all parallel code, with the major difference being one of scale.
The program spends some time computing and some time communicating, and
at a FIXED scale, especially a small one, one often finds that
communication times scale up to overwhelm computational advantages as a
problem is subdivided.  This is basically Amdahl's Law and its more
quantitative generalizations.

There is a whole chapter on this very question that derives at least
simple semi-quantitative scaling forms in:

  http://www.phy.duke.edu/brahma/beowulf_book/

or

  http://www.phy.duke.edu/~rgb/Beowulf/beowulf_book/

(which is the latest updated version -- I've actually been working on it
some once again).

You will very likely find that just making your code BIGGER will make it
scale well to ten nodes, so computation dominates communication.  This
is also explained in Sterling, Becker, et. al.'s book.  There are some
"talks" on the brahma site that go through parallelizing an actual
application and show how your experience is a perfectly normal one on
the first pass through.  

Finally, although a full analysis of your problem may well indicate that
you need gig ethernet or myrinet to get good scaling (if it is fine
grained and intrinsically has a lot of communication and is synchronous
and all that) you might ALSO look at a book or two on parallel
programming, as efficient parallel programming is not always intuitive
to serial programmers, even experienced ones.  Parallel algorithms are
often "different", and naive parallelizations of common tasks may be
very inefficient.  There is at least one online book on parallel
programming linked (IIRC) to brahma and referenced in my online book.

With a very limited number of nodes to scale to, there is no point in
getting faster communications unless you really need it, and you won't
know that until you try scaling up your application, studying your
parallel algorithm, and studying parallel programming in general.

> Now the question:
> What would be your advice about network cards (giga-ethernet seems
> clear but would they be fast enough?), and switches. My plan is not tu
> use more than ten computers in this first period of approach to beowulf.

We can't answer that or even help you answer it without much more
information.  You probably can't answer it yourself without learning a
bit more.  Study your code, work out its computation to communications
times (possibly with a profiler), see how the ratio changes when the
code is scaled up, see especially if it is latency dominated.  If
latency is killing you (lots of small packets) you may not get as much
of a boost moving to gigE as you might think.  Latency dominated
communications problems require high end networks to resolve, e.g.
myrinet as there can be an order of magnitude difference between
ethernet latencies and myrinet latencies (with other networks scattered
out across the field as well -- this isn't intended to disrespect any or
endorse any, and latency times can vary with implementation even within
a single paradigm).

If you look back at the list archives, there is an ongoing and lively
discussion on comparative virtues of the various networks, so getting a
definitive, comprehensive answer on which is "best" (has best
cost-benefit performance, meets your needs) for you will be VERY
difficult and will definitely require your active participation in
analyzing the details of your communications pattern.

   rgb

-- 
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



_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list