[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [HTCondor-users] Combining Accounting groups and Priority factors



On 04/24/2018 10:10 AM, Stephen Jones wrote:
Hi,

I have a question I've wondered about for quite a while.

Priority factors convert the usage of some user to a priority. A low priority factor will turn high usage into a low "effective priority" (where low is better). Hence users with a low priority factor will run more work. I think that's right. On the other hand, accounting groups assign users to a group, and an allocation can be made for that group, using (e.g. GROUP_QUOTA_DYNAMIC_group_SOMEGRP =Â 0.65)Â I've read somewhere that these can be used together. And looking at my site, I can see they _are_ used together. That's scary. What does it all mean?

The two ideas seem to be in conflict. One controls the allocation via a priority factor, that works on usage over time to promote (or penalise)Â certain users according to their relative recent usage. The other scheme "reserves" slots using GROUP_QUOTA_DYNAMIC... so which scheme (if any) takes precedence? And if none does, how does all this play out?

This is a good question, and at first blush, these two mechanisms seem to be in conflict.

The best way to think about this is that the group quotas "go first", and divide up one condor pool, usually as a result of some kind of ownership -- e.g. group a has purchased one third of the node, and group b has purchased the rest. After this subdivision, if groups have more than one user with jobs, HTCondor will still schedule jobs from those users in a "fair share" manner, with the decaying user priority.

-greg