You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
GCP recently made the dynamic workload scheduler feature available in public preview. This system works by letting you submit a request to get N number of GPU VMs and puts you in a queue to get them, only provisioning once all are ready. This increases the ability of users to get rare GPUs without needing reservations. It also decreases costs compared to grabbing 1 vm at a time until all are available.
It would be useful if when SkyPilot was provisioning GPU instances on GCE it created them through this system instead of standard node provisioning APIs.
I'd like to take this if possible. I'm trying to figure out where skypilot would fit in with Kueue currently :) I think this goes back to a discussion that was on the slack on how to make a batch job interface for k8s that treats Tasks/Jobs as first class objects instead of clusters.
colinjc
changed the title
[GCP] Add support for the Dynamic Workload Scheduler
[GCP] Add support for the Dynamic Workload Scheduler (GCE)
Apr 4, 2024
GCP recently made the dynamic workload scheduler feature available in public preview. This system works by letting you submit a request to get N number of GPU VMs and puts you in a queue to get them, only provisioning once all are ready. This increases the ability of users to get rare GPUs without needing reservations. It also decreases costs compared to grabbing 1 vm at a time until all are available.
It would be useful if when SkyPilot was provisioning GPU instances on GCE it created them through this system instead of standard node provisioning APIs.
Announcement - https://cloud.google.com/blog/products/compute/introducing-dynamic-workload-scheduler
GCE API - https://cloud.google.com/compute/docs/instance-groups/about-resize-requests-mig
The text was updated successfully, but these errors were encountered: