Hello SMEs,
I configured OpenPBS 20.0 for job allocations to GPU servers based on server load
Configured /var/spool/pbs/mom_priv/config like this->
$max_load 10.0
$ideal_load 5.0
This is working fine and allocation is happening based on CPU load of the GPu servers.
Can you please suggest how I can configure job allocations based on GPU usage rather than CPU load ?
For example: if there are 3 nodes , each with 2 GPU cards, and you would like to schedule a job on the node which has more unused GPU cards, then you can use node_sort_key against the ngpus (host consumable resource type long - resources_available.ngpus=XY)
For example: if you have an applicaiton that would report total load on the gpu load on the GPU-compute node and this value is updated to a custome resource value of compute node ( resources_available.gpuload=XYZ), then this value can be used in the node sort key and sorted based on the gpu load of the compute node.
Hope this helps.