[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [modeller_usage] parallel modeller





On Sun, Dec 7, 2008 at 12:53 PM, Mauricio Carrillo Tripp <">> wrote:
I ran into the same problem. I searched the wiki but couldn't find a solution. From the description
provided in the thread is not obvious how to proceed. Is there a working script that could be used
as a starting point to run parallel jobs in a cluster using PBS?

I managed to submit jobs using independent modeller.py scripts differing in the start and end model
parameter only:
in file modeller01.py start=1 end=1
in file modeller02.py start=2 end=2
each one of these run in a different node (jobs submitted with qsub), but the two resulting models
1s58.B99990001.pdb and 1s58.B99990002.pdb are exactly the same, which brings me to my second question:
Is there a way to specify the initial random seed used by each modeller script?


I was under the (wrong) impression that just by using different values for automodel.starting_model and
automodel.ending_model in two different modeller scripts (as described above) would produce two different
final models. The correct answer is that this will just produce the exact same final model but named differently.

I found automodel.rand_method=randomize.xyz and automodel.rand_method=randomize.dihedrals
(and I'm guessin automodel.rand_method=None being the default). If I add this declaration to the two modeller
scripts from above, would that have the same effect as to have only one script with automodel.starting_model=1
automodel.ending_model=2?



 


Thanks!


Starr Hazard wrote:
The references to parallelization seem to point rather strongly
to the SGE scheduler...
Not at all - the 'job' class is simply a bag of 'slave' objects. There is no requirement that you use any particular resource management system. For example, local_slave starts up a slave on the local machine (ideal if you have a multi-core machine). ssh_slave starts up a slave on a machine accessible by ssh, ideal if you have a cluster set up to allow passwordless ssh (or rsh) to individual nodes. The only slave classes which use SGE are sge_pe_slave and sge_qsub_slave. I wrote those because we happen to have an SGE cluster. But there's no reason why you couldn't write your own slaves to use PBS mechanisms.
Can any of the commands
eg

sge_qsub_job(options, maxslave, seq=(), modeller_path=None, host=None)

work with PBS scheduler?
Well, obviously not sge_qsub_job, as the name would suggest, since that is a convenience class for SGE. Just use the regular job base class. If you then have a traditional ssh-to-any-node setup, all you then need to do is loop over the nodes in your machine file and make an ssh_slave for each one. Alternatively, it would be pretty simple to write a class that used the PBS TM mechanism via something like mpiexec. I suggest you put the result into the Modeller wiki, so that other PBS users can use or modify it.



--
        0     |  Mauricio Carrillo Tripp, PhD
     /         |     Department of Molecular Biology, TPC6
 0            |     The Scripps Research Institute
     \         |     10550 North Torrey Pines Road
        0     |     La Jolla, California 92037
     /         |     " target="_blank">
 0            |  http://www.scripps.edu/~trippm

** Aut tace aut loquere meliora silentio **



--
        0     |  Mauricio Carrillo Tripp, PhD
     /         |     Department of Molecular Biology, TPC6
 0            |     The Scripps Research Institute
     \         |     10550 North Torrey Pines Road
        0     |     La Jolla, California 92037
     /         |     ">
 0            |  http://www.scripps.edu/~trippm

** Aut tace aut loquere meliora silentio **