StarCluster - Mailing List Archive

Re: GPU load sensor support

From: François-Michel L'Heureux <no email>
Date: Tue, 18 Dec 2012 11:40:15 -0500

Hi Mark!

This is more a question for the OGS mailing list (
users-request_at_gridengine.org) than the StarCluster one, but here is my
answer anyway.

I don't think it can be done by default. I would do it by adding your gpus
to complex variables. First, do

qconf -mc

There you may create your gpus as complex values. I would go with
gpu gpu INT <= YES YES {default?} 0

You can replace default with the number of gpus you want your requests to
take by default. Then, for each of your nodes, you must do

qconf -me {node name}

Then on the line "complex_values", you may add "gpu={num_gpus}". (Remove
the NONE string if it's present.) Of course, you have to replace "num_gpus"
with your count of gpus.

Then, when making your requests, you can add "-l gpu={num gpu requested}"
to your command. Taking your question as an example, you would do "qsub -l
gpu=2 ...". IMPORTANT: If you set 0 as default in qconf -mc, all your
requests using gpus must use the "-l gpu=" option, otherwise OGS will
assume they don't use any.

I hope this helps!
Mich


From: Mark Ebersole <markeber_at_gmail.com>
To: starcluster_at_MIT.EDU
Cc:
Date: Mon, 17 Dec 2012 15:27:59 -0700
Subject: [StarCluster] GPU load sensor support
Does the the Sun Grid Engine used in StarCluster support the use of load
sensors for the GPUs in a GPU-compute AMI? For example, can I do a qsub
and request a node(s) that has two free GPUs?

I searched through the features/bugs/mail archive, but don't see any info
on this.

Thanks!
Received on Tue Dec 18 2012 - 11:40:37 EST
This archive was generated by hypermail 2.3.0.

Search:

Sort all by:

Date

Month

Thread

Author

Subject