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

Re: [Condor-users] run a condor pool out of SGE nodes

On 11/14/2011 06:51 PM, Yu Huang wrote:
So I have a couple of idea to go around this problem. Before I go out
and try all of them, I want to see if someone has tried (searching
condor mailing lists turns up nothing or i don't optimize my keywords
hard enough).

 1. construct a condor pool out of SGE nodes by submitting jobs that run
    condor master/startd daemon in user space and custom ports. They are
    all managed by the one of the nodes, which acts as condor pool
    submission host. As i don't have root privilege whatsoever, I have
    run everything as normal user.

This is glideins:

The glideins are usually submitted over the GRAM interface, but as you want to avoid that, we can make a glidein which is submitted via the qsub command on the cluster login node.

It sounds like your plan is to have the central manager / submit node be one of the compute nodes. Do you have interactive access (ssh for example) to the compute nodes? Also consider that the environment is more important on the submit node, so it can be easier to have static set up for the central manager / submit host and just glidein the startds.

And if I recall correctly, your workload is a mix of memory requirements. We might want to have two types of glideins, one with the normal mem/N cores configuration and one with a high memory configuration. We could also test dynamic slots.

Let me know off-list if you want to meet and setup a test deployment.

 2. run the condor-g and globus in user space on the SGE cluster
    submission host. I have to configure the globus's GRAM part to talk
    to SGE.  And then I submit jobs on the cluster submission host.
    Again, everything has to be run in user-space.

This would just be a band aid, and probably not work any better than the current non-maintained GRAM install.

Mats Rynge
USC/ISI - Pegasus Team <http://pegasus.isi.edu>