Parallel MATLAB: Task Computing
3pm - 4pm, Friday, July 2 2010
3060 Torgersen Hall

Virginia Tech's new Ithaca cluster offers the research community
access to parallel MATLAB on up to 64 cores.  Interested users
need to learn about what kinds of problems can be handled in 
parallel, how one modifies an existing MATLAB program to express
a parallel algorithm, and how such a program is executed.

In this lecture, we will present the task computing model
of parallel computing with MATLAB.  We will concentrate on
one simplified version, what we call the "scattered task" model. 
In this case, a computation consists of a number of almost completely 
independent tasks, which can be carried out in any order, and which do not require
any communication or data sharing.  Only when all the tasks are
completed is there a chance to examine or gather the overall results.

This kind of programming is especially typical of Monte Carlo
calculations where, for instance, the computation consists of
running the same simulation many times, but each time with 
slightly different input.

We will show several example programs that work with this
task model.  We will cover the ways of running a parallel program 
on your desktop (interactively or indirectly), and how you can set 
up a link to Ithaca so that you can send jobs there and receive 
results without leaving your desktop MATLAB session.

MATLAB also allows for a more sophisticated kind of 
"parallel task programming", in which the tasks must run at the same 
time, and can communicate through messages.  This is one way to do 
systolic programming, for instance.  It is interesting to compare 
this with the usual way of doing this using SPMD commands.  However, 
we will probably be able to consider at most one example of this 
kind of parallel task programming during the lecture.

The slides for this class are available online at
"http://people.sc.fsu.edu/~burkardt/presentations/vt_2010_tasks.pdf".

Please note that, to make use of parallel MATLAB, your desktop
machine must have access to the Parallel Computing Toolbox.
This is automatically available as part of the campus-wide concurrent 
license version of MATLAB; however, it is NOT part of most standalone licences.

If you want to try out the direct desktop link to Ithaca, we encourage you to 
get an account on Ithaca in advance of the class, by going to 
http://www.arc.vt.edu/index.php, and, under "Services & Support"
selecting "User Accounts".  Once you have an account, you should
also refer to the document 
  "https://portal.arc.vt.edu/matlab/RemoteMatlabSubmission.pdf".
If you have trouble with the steps described in this document,
you may contact the instructors for advice and help.