Hi Stephen,
thanks for clarifying this. I'd be interested to know where the definition
is that you quote. Before mailing LCG-ROLLOUT, I searched through the LCG
admin guides and googled but to no avail.
Cheers,
Simon
On Mon, 24 Apr 2006, Burke, S (Stephen) wrote:
> LHC Computer Grid - Rollout
> > [mailto:[log in to unmask]] On Behalf Of Simon George
> said:
> > According to the yaim documentation:
> > CE_MINPHYSMEM :
> > RAM size in kblocks (WN specification).
> >
> > We have actually set ours to (RAM size of WN / no. of CPU in WN),
> > which is 1024/2 = 512MB.
>
> FWIW the definition says "configured physical memory on any one CPU in
> the subcluster in MB", which would tend to support what you've done, but
> as you observe it isn't entirely clear. In general the concept is to err
> on the safe side, so if you publish 512 and jobs request what they need
> (far from obvious!) then it should be safe, albeit jobs will sometimes
> be rejected when they could in fact run.
>
> > Even if a node ends up with 2 jobs needing a bit more physical memory
> > than there
> > is, there is always swap, which is fine as long as the job is
> > not actively accessing all the memory it uses.
>
> True, but presumably you would only specify such a requirement if it
> were fairly important for the job to get that much real memory.
>
> Stephen
>
|