we simply enforce --threads-per-core=2 in a job_submit plugin, which (in our case) ends up in having total_cpu=48 for each exclusive node. Hence, RawUsage is always given in CPU-seconds.
AFAIK there is no other usage of the option --threads-per-core than choosing a suitable node and "calculating" total_cpu. Would everyone agree on that?
On 09/01/15 14:21, Chris Samuel wrote:
On Fri, 9 Jan 2015 05:11:39 AM Hendryk Bockelmann wrote:Do you have any idea if this is an intended behaviour or a bug?We hit this same issue when we were bringing up our Intel Sandybridge cluster back in 2013 (as SMT for SB was better than previous generations) and it was a significant part of the reason for us disabling SMT, the pain it would cause us in managing quotas wasn't worth the gain. So we'd certainly prefer Slurm to account SMT's as being the same as their single core (or even better, to make that a configuration parameter so everyone can be happy). All the best, Chris
smime.p7s
Description: S/MIME Cryptographic Signature
