[HTCondor-users] Compute topology(?) questions and resource matching

[Matthew T West <m.t.west@xxxxxxxxxxxx>]

Afternoon All,

As vendors of x86 chips (arm?) continue to expand the core count on a 
single socket by using chiplets, I am curious if it is possible in job 
submit to request all the cores be on the same chiplet? For multi-core 
(single chip) workloads that need thread communication, performance 
would be better if they need not go over intra-die fabric.

On the complete other end of the size scale, I am curious how HTCondor's 
parallel universe handles jobs that need more machines that can fit on a 
single switch via 
https://htcondor.readthedocs.io/en/lts/admin-manual/setting-up-special-environments.html#grouping-dedicated-nodes-into-parallel-scheduling-groups. 
Is there some way to define "nearest neighbors" that folks who have run 
a dedicated scheduler have used in the past?

Cheers,
Matt

P.S. - sorry for getting all the HPC blah-blah on the HTC list.

--
Matthew T. West
DevOps & HPC SysAdmin
University of Exeter, Research IT
www.exeter.ac.uk/research/researchcomputing/support/researchit
57 Laver Building, North Park Road, Exeter, EX4 4QE, United Kingdom