[GE users] configure nodes to have a "head" node
dag at sonsorol.org
Mon Mar 20 13:25:57 GMT 2006
For a cluster of only 10 nodes it is extremely common for the head
node to serve double duty as both the user/submit node as well as the
machine that runs the SGE qmaster and schedd daemons. Most people
construct their cluster networks such that the compute nodes are not
accessible by end users except through the head node.
For your question #1 -- only you can really answer this since there
is no correct answer for everyone on how user actions influence the
performance of SGE qmaster and scheddd. It depends on how many active/
pending jobs you may have entering and exiting the system. For a
small cluster doing computational biology and bioinformatics the user
induced load is probably not going to be all that huge. Bio people
tend to run lots of IO intensive serial jobs rather than slamming SGE
with hundreds of thousands of small parallel tasks. Your biggest
performance bottleneck will be storage and fileservers if you are
doing bioinformatics on this system.
If you want to start experimenting with #1 consider adding an SGE
queue instance with only a single job slot on it - see how the system
handles its tasks along with a single grid job running.
For question #2 -- the mechanism you would use within Grid Engine to
manually influence the order in which machines are picked is called
"queue_sort" -- you can assign each of your queue instances an
integer value and tell SGE to sort first on the queue number and then
on load. The end result is you can control the order of machine
dispatch and you can place your head node at the "back" of the line.
There are a few blog entries on queue sorting and you can also search
the mailing lists for when these methods have been discussed:
On Mar 20, 2006, at 3:32 AM, Davide Cittaro wrote:
> Hi all, I'm pretty new to sge6. I have this 10 nodes cluster I set
> up with sge6. One node is only for qmaster and qsched, there is no
> user access, no work at all. The other 9 nodes are submit/execution
> Now I have a problem. Some users would like that their works do not
> migrate on node1 (the node tipically used to submit works) except
> when the other 8 nodes are really full. In other words they would
> like to have a "head" node.
> I'm thinking to two solutions and I need help for both:
> 1- let the qmaster node be used as the unique submit host. In this
> case I would like to know how much the users activity on the node
> influences the efficiency of the qmaster and qsched daemons...
> 2- let node1 being considered less "performant" or "important" so
> that sge chooses to migrate works on nodes2-9 and, in the end,
> node1. In this case I would like to know which parameters I can
> tune. Consider that all 10 machines are hardware identical (2
> opteron, 4Gb RAM).
> Davide Cittaro
> Bioinformatics Systems @ Informatics Core
> IFOM - Istituto FIRC di Oncologia Molecolare
> via adamello, 16
> 20139 Milano
> tel.: +39(02)574303355
> e-mail: davide.cittaro at ifom-ieo-campus.it
To unsubscribe, e-mail: users-unsubscribe at gridengine.sunsource.net
For additional commands, e-mail: users-help at gridengine.sunsource.net
More information about the gridengine-users