So what if you set the default numprocs on a cluster to the max ppn? (I think it is that way already). Would the only problem then be doing the 5 proc jobs? Just identifying the need here...
Jeremy
At 03:52 PM 10/5/2001 -0500, Jim Basney wrote:
My goal is to submit a job that requires N processors (where N is less than or equal to the number of processors in the cluster), without needing to know if the cluster consists of single-processor machines or SMP machines. I want PBS (or Maui) to figure out how to give me N processors in the cluster. For example, I don't want to have to tell PBS that I want 4 dual-processor machines or 8 single-processor machines (or 2 dual-processor machines and 4 single-processor machines) -- I just want 8 processors. MPI will choose shared memory or network communication for me as appropriate once I get my 8 processors.
I'm the sysadmin for the different clusters. I can make any configuration changes I need to accomplish this goal but I want my users to use the same job submission script on all of my different clusters without needing to know the specific machine configuration of each cluster.
-Jim
At 03:32 PM 10/5/2001, Jeremy Enos wrote:
I'm not sure that PBS supports multiple ppn specifications within the same submit. Also, keep in mind that some admins might to post configuration that changes PBS's default ppn to 1, thus preventing you from filling the cluster with that command at all.
There is one other way to do it... but it's kind of a hack. qsub will take switches to make a job not start until after another job starts, after or another job finishes, or also exactly when another job starts (I think). For this, you would have to submit multiple jobs to make up the same one though, and that is way too much of a kludge if you ask me.
Is your goal to be able to find out the max number of procs available for a job, or to be able to launch the same job w/ different PPN specifications?
Jeremy
At 11:59 AM 10/5/2001 -0500, Jim Basney wrote:
Hello,
Jeremy Enos and I installed OSCAR 1.1 on a cluster of 5 dual-processor machines (1 head node and 4 compute nodes) here at NCSA for our Grid-in-a-Box project. I'm hoping you can answer (what I hope are) some simple PBS questions for me. I want to be able to submit mpich jobs to OSCAR clusters without needing to know how many processors each machine in each cluster has. Instead, I just want to know how many total CPUs each cluster has. This simplifies my life when doing "Grid computing", where I'm submitting to different clusters across the network, where some clusters may have dual-processor machines and others may have single-processor machines.
When I submit an MPI job to PBS on our cluster with "qsub -l nodes=4", PBS allocates 2 processors each on 2 compute nodes to my job, as opposed to allocating 1 processor each on 4 compute nodes. That's great because it makes me think "nodes" in PBS terminology means CPUs rather than machines. However, when I try "qsub -l nodes=8", I get "Job exceeds queue resource limits".
Is there some trick I can play to get this to work without needing to set ppn=2 in my qsub line on the clusters with dual-processor machines? If I have to use ppn, how would I request 5 CPUs from the cluster? "qsub -l nodes=2:ppn=2+nodes=1:ppn=1" fails with "Job exceeds queue resource limits" but "qsub -l nodes=3:ppn=2" succeeds. Is my syntax wrong?
Any help would be greatly appreciated.
-Jim
_______________________________________________
Oscar-users mailing list
[EMAIL PROTECTED]
https://lists.sourceforge.net/lists/listinfo/oscar-users
_______________________________________________ Oscar-users mailing list [EMAIL PROTECTED] https://lists.sourceforge.net/lists/listinfo/oscar-users_______________________________________________ Oscar-users mailing list [EMAIL PROTECTED] https://lists.sourceforge.net/lists/listinfo/oscar-users
_______________________________________________ Oscar-users mailing list [EMAIL PROTECTED] https://lists.sourceforge.net/lists/listinfo/oscar-users </x-flowed>
