Aaron, You have discovered a case where the design/implementation outpaced the documentation. Originally the usage was normalized to the theoretical maximum usage subject to half-life decay as documented. But we found that, for under-utilized systems, each user's normalized usage was very low, which gave them higher fair-share factors. In this circumstance, the result was that fair-share factors for all users tended to get crowded toward 1.0 and one would have to increase the fair-share weight to resolve them.
By changing to make the usage normalized to total actual usage, we prevented this artifact. It made the code a little simpler as well. With this change, we still maintain "fairness". However, users may notice minor variations in the normalized usage that gets reported based on usage outside their accounts. This variation will diminish substantially the more fully a cluster is utilized or by increasing the PriorityDecayHalfLife configuration value. I will update the html page to reflect the newer formula. Don From: [email protected] [mailto:[email protected]] On Behalf Of Aaron Knister Sent: Wednesday, May 25, 2011 7:22 PM To: slurm-dev Subject: [slurm-dev] Normalized usage question The multifactor priority documentation seems to suggest that the normalized usage is calculated based on the cluster's available CPU time. However, in practice it seems to be based on the summation of all raw usage. I'm getting ready to implement fairshare priorities and am wondering which is the case. The former would be ideal as that way the fairshare values in the output of sshare won't flip flop based on usage outside of a given account. Of course, it's always possible something is borked in my testing setup :)
