RE: Embarrassingly Parallel Tasks on HPC

[Kokron, Daniel S. (GSFC-610.1)[Computer Sciences Corporation]]

Erik,

Does the carver system 'require' interactive password upon connection to PBS 
assigned nodes as a matter of administrative policy?

Daniel Kokron
NASA Ames (ARC-TN)
SciCon group
301-286-3959

________________________________________
From: parallel-bounces+daniel.s.kokron=nasa.gov@gnu.org 
[parallel-bounces+daniel.s.kokron=nasa.gov@gnu.org] On Behalf Of Erik Petigura 
[eptune@gmail.com]
Sent: Wednesday, August 29, 2012 1:10 PM
To: parallel@gnu.org
Subject: Embarrassingly Parallel Tasks on HPC

I'm working on a project to search for transiting planets in 10^4 light curves 
collected by the Kepler space telescope.  The search is implemented using a 
series of ~5 python scripts in the following manner.

python script1.py star 1
<...>
python script1.py star N
python script2.py star 1
<...>
python script2.py star N

and so on.  N~10^4 and total CPU time required is 3000 hours.  I've had a great 
deal of success implementing this with gnu parallel on a small cluster that 
uses PBS for resource management.

cat cmdList.txt | parallel --slf $PBS_NODEFILE

Now, I'm trying to migrate this code to a larger computer ("carver" at NERSC).  
The command no longer works because carver requires a password for each ssh 
connection to a child node.  It seems like MPI is the preferred method for 
communicating between nodes, but since this work is embarrassingly parallel, 
MPI seems like overkill.

Does any one have any advice?

Erik