Hi Sam,
On 2017-04-14 10:01, Sam Skipsey wrote:
> the recommendations from the sites, and scheduling experts, were to
> increase
> the number of jobs which can be pre-empted, try not to schedule really
> wide
> jobs, and, if possible, to keep a mix of 'multicore' and 'single-core'
> jobs
> in the queue.
Yes to much of that. Perhaps the sticking point today is "jobs which can
be pre-empted." We have to signal that, I think (unless we already do?)
> Ideally, of course, WLCG jobs would remove the requirement that
> multicore
> jobs have all processes scheduled on a single node.
In my view, as far as I can tell (which is not _that_ far because I'm an
old slow guy) the usefulness of "multicore" is that the cache is filled
by the first thread, and thus less loading and less ram is needed by the
subsequent threads (because the cache is not loaded further.) Hence,
right now, multicore jobs must reside on one node. It's an unfortunate
(or fortunate?) architectural hardware constraint (delete as
appropriate).
Cheers,
Ste
|