[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [HTCondor-users] Rank with Partitionable and Dynamic Slots
- Date: Wed, 27 Aug 2014 09:15:09 +0200
- From: Steffen Grunewald <Steffen.Grunewald@xxxxxxxxxx>
- Subject: Re: [HTCondor-users] Rank with Partitionable and Dynamic Slots
On Tue, Aug 26, 2014 at 11:03:23AM +0200, matthias.roehm@xxxxxxxxxx wrote:
> Adding the following to the machine config file did the trick:
> RANK = $(RANK)
As that look quite tautological at first and second sight, I'd love to
learn how it actually works... and whether that would help me to fix my
I've been struggling with my own ranking of dynamic/part'able slots,
trying to hand out partially consumed machines first, with
NEGOTIATOR_PRE_JOB_RANK = 1000000000 - 10000000 * Cpus - 1000 * Memory (...)
(there are a few more terms in that expression which don't matter for now).
This should favour smaller "remainders" over "full machines", in theory.
I'd expect jobs to cluster on a small number of machines (that have already
been partitioned, or even re-using dynamic slots still in existence), but
what I see is a rather random distribution, not even strictly leaving out
the couple of machines which come with bigger memory and more CPU cores.
Do these observations coincide with the RANK issue, and how to fix the
expression so it actually does what I intended it to do?