MATLAB should not be
run on the SciClone front end node (monsoon)
unless you are doing something truly trivial. For operations which use
significant amounts of CPU time or memory, use the qlogin
command to obtain an interactive shell on a compute node, or else run
MATLAB in batch mode from a PBS job script. Note that if you use qlogin
to start an interactive PBS job, the MATLAB graphical interface will
still work properly assuming that:
- the system you logged
in from is running an X11 server,
- your local SSH configuration
supports X11 forwarding, and
- your network connection
to SciClone is fast enough to support a remote X session.
Additional information
about establishing X11 connections to remote workstations can be found
in the SciClone
User's Guide.
To run MATLAB non-interactively
as a PBS batch job, use the "-nojvm" and "-nodisplay"
options and read your command script from standard input, e.g.:
qsub -l nodes=1:c7:ppn=1
#!/bin/csh
cd mydir
matlab -nojvm -nodisplay <script.m >results.out
^D
Multiple copies of MATLAB
can be invoked independently on different nodes and/or on multiple processors
within a single node to exploit coarse-grained parallelism. Because
the College has purchased a site license with essentially unlimited
usage, the maximum number of simultaneous MATLAB jobs is limited only
by the number of processors available in the system and the job
queue parameters.
Interactive MATLAB sessions
require large amounts of memory (minimum of 1 GB is recommended) and
will not run on SciClone's c1,
c2, or c6 compute nodes. Non-interactive jobs (without the desktop
GUI and Java) use much less memory and will run on any node type, subject
to the application's memory requirements.
Despite the fact that
Solaris has been 64-bit capable since 1997, MathWorks has yet to release
a 64-bit version of MATLAB for Solaris. This means that MATLAB applications
on SciClone are currently limited to 2 GB of memory. We have asked MathWorks
to rectify this situation, and are hopeful that a 64-bit Solaris version
will become available at some point in the future.
MATLAB is not presently
available for x86/x64 systems running Solaris, and therefore is not
installed on SciClone's tempest subcluster.
MATLAB 7 introduced a
capability for distributing computations across multiple processors
via the Distributed Computing Toolbox. This feature has not been
integrated into SciClone's PBS batch environment, although it is probably
possible to do so. If you are interested in using the Distributed Computing
Toolbox on SciClone, please contact sciclone@compsci.wm.edu
for more information.
