What you would like to be added?
Grove should expose an API for requesting preferred topology constraint at every level:
- PodClique
- PodGangScalingGroup
- PodCliqueSet
Currently only required is available via packDomain param
Why is this needed?
Required might fail the workload completely, while at some scenarios user will only want best effort even if placement will eventually be sub-optimal without failing the workload. It also compatible with other schedulers such as Kueue.
The reason preferred can't apply by default by the scheduler and requires the user to request it - is because it is very compute intensive algorithm.
Specifically for the flow of submission of Dynamo over Grove from Run:ai, this will help as the system automatically applies preferred constraints to all the workloads submitted, to abstract this from the end users.
What you would like to be added?
Grove should expose an API for requesting preferred topology constraint at every level:
Currently only required is available via
packDomainparamWhy is this needed?
Required might fail the workload completely, while at some scenarios user will only want best effort even if placement will eventually be sub-optimal without failing the workload. It also compatible with other schedulers such as Kueue.
The reason preferred can't apply by default by the scheduler and requires the user to request it - is because it is very compute intensive algorithm.
Specifically for the flow of submission of Dynamo over Grove from Run:ai, this will help as the system automatically applies preferred constraints to all the workloads submitted, to abstract this from the end users.