Chris
I am unfamiliar with the details of SGE; I cannot tell whether this approach makes sense.
-erik
On Thu, Oct 7, 2021 at 5:19 PM Chris Stevens chris.stevens@canterbury.ac.nz wrote:
Hi Erik,
Thanks for your suggestion.
I am happy using these in the scripts, but I think the problem is how to pass these expressions to SGE. From what I can tell, the output of @(@PPN_USED@/@NUM_THREADS@)@way is, for example, "6way", given @PPN_USED@=48 and @NUM_THREADS@=8. This means that I have requested the parallel environment called 6way with @PROCS_REQUESTED@ slots. If I requested 48 slots, then I would use mpirun -np 6. Thus, from what I gather, for this to work, this specific parallel environment 6way needs to exist. I am now figuring out how to configure parallel environments in such a way, most likely by changing the allocation rule.
Let me know if you think this is wrong, as it does seem rather stupid to not be able to just set -ncpus-per-task like in Slurm in the submission script.
Cheers,
Chris
*Dr Chris Stevens*
*Lecturer in Applied Mathematics*
Rm 602, Jack Erskine building
School of Mathematics and Statistics
T: +64 3 369 0396 (Internal 90396)
University of Canterbury | Te Whare Wānanga o Waitaha
Private Bag 4800, Christchurch 8140, New Zealand
*Director* SCRI Ltd http://www.scri.co.nz
*From:* Erik Schnetter schnetter@cct.lsu.edu *Sent:* 08 October 2021 09:40 *To:* Chris Stevens chris.stevens@canterbury.ac.nz *Cc:* users@einsteintoolkit.org users@einsteintoolkit.org *Subject:* Re: [Users] Einstein toolkit with Sun Grid Engine
Chris
It might not be necessary to hard-code the number of threads. You can use arbitrary Python expressions via "@( ... )@" in the templates. See e.g. the template for Blue Waters which uses this to choose between CPU and GPU queues.
-erik
On Thu, Oct 7, 2021 at 4:04 PM Chris Stevens < chris.stevens@canterbury.ac.nz> wrote:
Hi Roland,
That's fantastic, thanks for linking those files.
It works as expected with only MPI processes. I am careful in compiling and running with the same (and only) OpenMPI installation on the cluster, so this should be OK.
Finding a Slurm to SGE conversion table, there is no SGE equivalent to ncpus-per-task from Slurm, rather it is the allocation type of the given parallel environment that does this. I.e. the backend.
https://srcc.stanford.edu/sge-slurm-conversion
Further, in the submit script of ranger, the crucial line
#$ -pe @(@PPN_USED@/@NUM_THREADS@)@way @PROCS_REQUESTED@
shows that you request @PROCS_REQUESTED@ slots (as I currently have) and the first argument shows that the name of the parallel environment is dependent upon @NUM_THREADS@. From what I take from this, I need to set up a parallel environment that has hardcoded the number of threads I want per MPI process and then use that parallel environment. I'll see how I go there, but it isn't initially obvious how to do this!
Cheers,
Chris
*Dr Chris Stevens*
*Lecturer in Applied Mathematics*
Rm 602, Jack Erskine building
School of Mathematics and Statistics
T: +64 3 369 0396 (Internal 90396)
University of Canterbury | Te Whare Wānanga o Waitaha
Private Bag 4800, Christchurch 8140, New Zealand
*Director* SCRI Ltd http://www.scri.co.nz
*From:* Roland Haas *Sent:* Thursday, October 07, 2021 06:22 *To:* Chris Stevens *Cc:* users@einsteintoolkit.org *Subject:* Re: [Users] Einstein toolkit with Sun Grid Engine
Hello Chris,
We used SGE a long time ago on some of the TACC machines.
You can find an old setup for TACC's Ranger cluster in an old commit like so:
git checkout fed9f8d6fae4c52ed2d0a688fcc99e51b94e608e
and then look at the "ranger" files in OUTDATED subdirectories of machines, runscripts, submitscripts.
Having all MPI ranks on a single node might also be caused by using different MPI stacks when compiling and when running so you must make sure that the "mpirun" (or equivalent command) you use is the one that belongs to the MPI library that you used when linking your code.
Finally you may also have to check if this is an issue with threads and MPI ranks. Ie I would check if things are still wrong if you use only MPI processes and no OpenMP threads at all (in that case you would have to check what SGE counts: threads (cores) or MPI ranks (processes)).
Yours, Roland
Hi everyone,
I have set up the Einstein toolkit on a local cluster of 20 nodes with
the SGE scheduler. I have not seen any examples of this scheduler being used with the Einstein toolkit.
I have managed to get it working; however it seems if I ask for a
certain number of slots that requires more than one node, it correctly allocates these, however all processes and threads are run on the one node and is oversubscribed.
My question is whether anybody has used SGE with the Einstein toolkit
and if this is a good thing or not? If it is possible, I can send more details if there are people willing to help solve this inter-node communication problem.
Thanks in advance,
Chris
[cid:29d54967-59c8-486e-adea-80af7ce2cc49]
[cid:55ebbbb5-1e12-45a2-8d51-206c70460c36]
Dr Chris Stevens
Lecturer in Applied Mathematics
Rm 602, Jack Erskine building
School of Mathematics and Statistics
T: +64 3 369 0396 (Internal 90396)
University of Canterbury | Te Whare Wānanga o Waitaha
Private Bag 4800, Christchurch 8140, New Zealand
https://urldefense.com/v3/__http://www.chrisdoesmaths.com__;!!DZ3fjg!rvExVfo... < https://urldefense.com/v3/__http://www.chrisdoesmaths.com/__;!!DZ3fjg!rvExVf...
Director SCRI Ltd
https://urldefense.com/v3/__http://www.scri.co.nz__;!!DZ3fjg!rvExVfoK3iWdskf... < https://urldefense.com/v3/__http://www.scri.co.nz/__;!!DZ3fjg!rvExVfoK3iWdsk...
-- My email is as private as my paper mail. I therefore support encrypting and signing email messages. Get my PGP key from http://pgp.mit.edu . _______________________________________________ Users mailing list Users@einsteintoolkit.org http://lists.einsteintoolkit.org/mailman/listinfo/users
-- Erik Schnetter schnetter@cct.lsu.edu http://www.perimeterinstitute.ca/personal/eschnetter/
Users mailing list Users@einsteintoolkit.org http://lists.einsteintoolkit.org/mailman/listinfo/users
