I could start by asserting
that if you have a REAL supercomputer,
you are not worried about RAM ;)
If it does become an issue, you could try
using a smaller number of gfServer instances
mixed with a somewhat larger number of gfClient
instances. The gfClient still needs to use
ram for the alignments that it is working on,
and does a significant amount of work.
We don't have anything that does this
automatically.
As Hiram mentioned, depending on what
you are doing you might want to split
up the query and target as another way
to both parallelize your job and
possibly reduce resource requirements.
Around here we are mostly running on
Beowulf-style commodity clusters
with a few hundred machines.
They tend to have 2 to 4 CPUs or cores,
and we tend to run an equal number
of BLAT jobs on each with Parasol.
Although not optimized for use on a supercomputer specifically,
Parasol is happy to run a number of jobs
on the machine. Simply supply the appropriate
config file. However, you may use any system that you like for running
your BLAT jobs.
-Galt
Ar 7/5/2010 12:27 PM, scr�obh Assaf Gordon: