[Beowulf] Grid Engine question

[Sean Dilda]

On Tue, 2004-10-12 at 04:11, Mark Westwood wrote:
> Hi All
> 
> We use the open source Grid Engine, Enterprise Edition v5.3, here to 
> manage job submission to our 70 processor Beowulf.  I'm rather new to 
> managing Grid Engine and my users have me baffled with a question of 
> priorities.
> 
> The scenario is this:
> 
> - suppose that there is a job running on 40 processors, leaving 30 free;
> - a high priority job, requesting 64 processors, is submitted;
> - a low priority, but long, job, requesting 24 processors is submitted.
> 
> Currently, with our configuration, the low priority job would be run 
> immediately, since there are more than 24 processors available. 
> However, my users want to hold that job until the high priority job has run.
> 
> Can we configure Grid Engine so that the low priority job is not started 
> until after the high priority job, even though there are resources 
> available for the low priority job when it is submitted ?

In SGE 6.0 they added a feature they call 'advanced reservations'.  Its
not really advanced, and its not what I consider 'reservations' to be,
but it is exactly what you want.  When reservations are enabled on the
cluster, and the job is submitted with '-R y', the mutli-processor job
will be able to 'hold' available resources until it has enough to run,
and thus keep lower priority jobs from using them.

However, to do this you need to upgrade to at least version 6.0. 
However, 6.0 also has cluster queues which I find makes administration
much easier (it allows you to create one queue setup and assign it to
multiple hosts instead of doing a separate setup for each compute host).