-
Notifications
You must be signed in to change notification settings - Fork 134
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Better fairness metrics #3789
Comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
(definition: blocked user = user that has jobs queued that could fit onto an empty farm but weren't put on in the most recent scheduling round)
(1) For "resource" in [CPU, memory, disk, sum-of-fractions-of-capacity]: What fraction of resource does each blocked user have, relative to priority_thisuser / sum(priority among blocked users) (where large priority = more jobs). "Perfect fairness" would be that each user's line plot hovers around 1 (with each user's line cutting out when they're not currently blocked)
(2) For "resource" in [CPU, memory, disk, sum-of-fractions-of-capacity]: What's the smallest job (including armada-id) that didn't get on in the last round of scheduling for each blocked user (for gangs, maybe just sum across jobs-in-gang?). If previous plot is not "perfectly fair", people would probably be less likely to complain if they asked for (32core, 500GBram) jobs while top user asked for (1core, 16GB) jobs (although I personally still think the scheduler should be improved to try harder to give big jobs equal share, it's at least understandable that that's a non-trivial problem)
(3) For "resource" in [CPU, memory, disk, sum-of-fractions-of-capacity]: Plot of cumulative (from midnight each day) preempted resource-times-runtimeAtPointOfPreemption; by-user-and-stacked-over-all-users
The text was updated successfully, but these errors were encountered: