Re: [HTCondor-users] assigning multiple GPUs to a single slot

[John M Knoeller <johnkn@xxxxxxxxxxx>]

Since you are not using partitionable slots, the GPUs will be permanently assigned to slots when the STARTD starts.

If you have only 2 gpus and want 2 slots, Each slot will be assigned only a single GPU â so jobs that want more than

1 GPU will not match any of your slots. 

 

You can instead both GPUs to 1 of your two slots so that slot will match jobs that want 1 or 2 GPUs.

 

SLOT_TYPE_1 = cpus=1, GPUs=2, mem=auto

NUM_SLOTS_TYPE_1 = 1

SLOT_TYPE_2 = cpus=1, mem=auto

NUM_SLOTS_TYPE_2 = 1

 

Or you can switch to using partitionable slots, and let HTCondor decide how to divide up resources based on

What the jobs request.  Be aware that if you do this, the 1 GPU jobs will tend to dominate (if you have an infinite

supply of them), since once a 1 GPU job starts the remainder of the partitionable slot will only match 1 GPU jobs.

 

-tj

 

From: HTCondor-users [mailto:htcondor-users-bounces@xxxxxxxxxxx] On Behalf Of Francisco Pereira
Sent: Monday, January 18, 2016 1:39 PM
To: Condor-Users Mail List <condor-users@xxxxxxxxxxx>
Subject: [HTCondor-users] assigning multiple GPUs to a single slot

 

Hi,

 

I have been scheduling GPU jobs in our cluster by 

 

1) setting in the config file for each node

 

"use feature: GPUs"

"GPU_DISCOVERY_EXTRA = -extra"

 

(as suggested in the documentation, running condor_gpu_discovery -properties manually produces the right results for each machine)

 

2) setting up a number of slots with 1 CPU each, e.g. in a 2-GPU machine.

 

"SLOT_TYPE_1 = cpus=1,mem=auto

SLOT_TYPE_1_PARTITIONABLE = FALSE

NUM_SLOTS_TYPE_1 = 2"

 

When submitting jobs that have "request_GPUs=1" in the submit file the jobs get scheduled to machines that have a GPU, and there are no more jobs being scheduled than there are GPUs, across multiple machines. However, when I specify "request_GPUs=2", the job stays in the queue with status "I", even though the requested number is available.

 

Hence, I am wondering what I am doing wrong and whether I have incorrectly set up the basic mechanism in #2. The GPU discovery works beautifully, so I suspect I am overcomplicating ... 

 

thank you for your help!

Francisco