> -----Original Message-----
> From: Testbed Support for GridPP member institutes [mailto:TB-
> [log in to unmask]] On Behalf Of Alessandra Forti
>
> ATLAS is moving to use RSS values to broker the jobs. I'll do the first
> sweep of changes assuming 2GB/job slot. If you have/want a different
> value please let me know.
>
I think the answer is somewhat dependent on what the jobs actually need - the most recent Oxford worker nodes have 4GB of memory per physical core, but we turned the hyperthreading on and set them up as 2GB per job slot; I think that was a fairly common pattern.
If the jobs are all completely happy with 2GB each then we can stick with that, but the plan was always to reduce the number of slots per node as the memory requirements went up, so it's slightly less a question of what the site wants, and more whether the VO would prefer to trade off job slots for more memory.
(also, a thought to add to the recent accounting discussion - should slots be accounted something like benchmark*time*RAM in the future?)
Ewan
|