[GE users] Help on designing simple queues

Lönroth Erik erik.lonroth at scania.com
Sun May 13 14:28:18 BST 2007

    [ The following text is in the "iso-8859-1" character set. ]
    [ Your display is set for the "ISO-8859-10" character set.  ]
    [ Some special characters may be displayed incorrectly. ]

My first post here, so bare with me.
I have a single SGE cell with 70 nodes. Each node has 4 CPUs. (2 x 2)
I have 2 separate HSI (myrinet mx) using 35 + 35 nodes. The HSIs are separeted, so submitted jobs need to be running on a specific HSI partition for performance.
I'm refering to the two partitions as:
Partition:         [ts102]   [ts103]
Nodes x CPU:         [35x4]    [35x4]
In summary there are 2 "subclusters" with regard to the HSI:s. I want this to be transparent to our users (who sees this a single cluster cell) and to get SGE secure that jobs are never split between the two "subclusters". I also never want more than 4 "job-processes" running on any node at a time.
Question 1: How can I setup on a best practice, so that a single job never cross the two HIS:s (subclusers) and the user wont have to change jobscripts or anything between submitts. (I dont want to specify different -p mpi_x ).
Question 2: How can I setup my queues so that a single parallell job wont be split over two different ques once it has got assigned one? (This happens in my lab-sessions alot)
I have started a setup of 4 queues at the moment (ts102 and ts103 are my subclusters):   
All queues contain all my nodes.
All queues should be able to run all applications.
All queues have all parallell environments.
I'm not afraid of ripping this apart to enforce a "best practice" for my environment.
I realize I'm a beginner here and I apprechiate all help and pointers to reading on this and hopefully you can assist me in deriving a good strategy on setting up my environment. We run some 5-6 different application on the cluster and we need to have a maintainable situation with not too many queues to maintain. SGE has so far impressed us at Scania and we will probably expand our usage of it as we learn.
Best regards.
/Erik Lonroth - Scania Infomate AB - Sodertalje, Sweden

More information about the gridengine-users mailing list