[glue-wg] Updated thoughts...

[Burke, S (Stephen)]

Paul Millar [mailto:paul.millar at desy.de] said:
> Maarten.Litmaath at cern.ch said:
> [StorageEnvironment RP multiplicity]
> > I argued that GLUE probably should allow for the RP being 
> multi-valued,
> > and probably the AccessLatency as well.  In WLCG/EGEE we 
> would have a
> > single value for each normally, so that the Storage Class is clear.
> 
> I'm not sure what AccessLatency as a multivalue means: I'd 
> push this attribute 
> down to the hardware layer 
> (StorageDatastore/StorageMedia/StorageStorage)

Still going through old mails ... I believe the decision on this was to
allow RP to be multivalued, even if not with the current technology, but
not AL. Re the discussion over the relational implementation we should
remember that that has a big overhead for multivalues, particularly if
you may want to select on them, so we should probably try not to
introduce them too freely. I can just about imagine that RP could be
dynamic, e.g. for custodial the system could make more copies, so maybe
it is worth allowing for that, but I still think that AL (and Tag)
should be single-valued.

> It's possible that we've been talking slightly at 
> cross-purposes.  It seems to 
> me that what WLCG means by "access latency" is really the 
> minimum (ie, fastest) guaranteed access latency (MGAL):

Yes - we had something of a discussion about this and tried to improve
the wording, maybe you can see if it helped. As you observe part of the
problem is that with latency smaller is better, so "minimum" sounds bad
but is actually good. However it isn't LCG that defines it, AIUI that's
the general SRM definition. When you're actually reading or writing a
file it's always online ... also files can be pinned online but that
doesn't count as "guaranteed" for these purposes, pinned files are still
just MGAL nearline.

> AccessLatency is a property of the hardware,

Strictly speaking that isn't always true, a file on disk may still get
copied to a different disk before becoming readable and that can take a
non-negligible time, but it probably isn't possible to take account of
that.

> I guess this is: an StorageEnvironment is 
> considered custodial if files are always stored within at least one 
> StorageMedia/Datastore that is considered custodial.

Not necessarily; you might just implement custodial by making multiple
copies (which might even be more secure than tape - tapes do get
corrupted after all). It's just LCG that insists on tying it to the
hardware.

Stephen