
Hi,
What do folks think about changing the default MODIFY_REQUEST_EXPR_REQUESTMEMORY?
Looking at our current cluster, it's mostly just doing damage, as it can greatly increase the job's memory request. See below - I don't feel like adding up all the memory rounding, but the last two rows correspond to 800GB of RAM wasted.
The worst case scenario is the default wastes 25% of throughput. In practice, we sometimes come close to this. No efficiencies gained in the negotiator is worth that.
I recommend changing the default to:
MODIFY_REQUEST_EXPR_REQUESTMEMORY = quantize(RequestMemory, 100)
Brian
> condor_q -g -format "%d\n" RequestMemory | sort | uniq -c | sort -n
4882 1900
6882 2500
> condor_status -const 'DynamicSlot=?=TRUE' -format "%d\n" Memory | sort | uniq -c | sort -n
10 2676
66 2415
81 2376
100 2250
132 2970
134 2392
210 2404
294 2990
360 2700
586 3005
702 3220
Attachment:
smime.p7s
Description: S/MIME cryptographic signature