[HTCondor-users] Defrag finer than whole machine granularity

["Fischer, Max (SCC)" <max.fischer@xxxxxxx>]

Hi all,

we are having problems formulating an ideal DEFRAG policy as we do not need entire nodes as "whole", but smaller fractions. We need roughly 50% of our cluster to be multicore, but it is not a fixed fraction - static allocation would not work.

Our biggest required slot is 8 cores, but our machines are 24-48+ cores.
If we consider a whole machine as having a single 8 core slot, then only 15%-30% of the cluster are multicore.
If we consider a whole machine as having a n*8 slot care slot, then DEFRAG also drains 8 core slots.

- Is there a way to express a DEFRAG policy to desire multiple 8 core slots (i.e. 2 8 core slots instead of 1 16 core slot)?
I see that the slots advertise e.g. `ChildCpus = { 1,1,1,1,8,1,1,1,1 }`, but do not know how it is possible to desire e.g. two 8 core slots to be in there. I've only found the reference for old-style classads [1].

- Is there a way to have DEFRAG_MAX_WHOLE_MACHINES reflect the number of e.g. 8 core slots instead of whole machines?
We can easily derive how many 8 core slots we need (at most), but the heterogeneous resource pool means that this does not translate to a fixed number of machines.

Cheers,
Max

[1]
http://research.cs.wisc.edu/htcondor/manual/current/4_1HTCondor_s_ClassAd.html

Attachment: smime.p7s
Description: S/MIME cryptographic signature