LHC Computer Grid - Rollout
> [mailto:[log in to unmask]] On Behalf Of Jeff Templon
said:
> Seems to me the concept of 'close' SE has been problematic
> since the BOG
> (Beginning Of Grid). If we ever agree on a definition of
> 'close' -- and
> I doubt it, just try discussing what a 'dataset' is with somebody --
> then we could use it.
It has traditionally meant at least three different things: 1) a default
SE to use for writing files if no destination is explicitly specified;
2) an SE you can use for reading files from a WN from which the access
can be expected to be "fast" in some undefined sense; 3) an SE to which
you can get "local" access from a WN for protocols like NFS and rfio
which only work within a site.
The first of those has been superseded by a VO-dependent environment
variable in LCG for some time, and that should now be explicitly
published in the new Glue schema. The third case was never very explicit
and didn't work very well; NFS has been out of use for some time and
rfio is not much used so it hasn't been that much of a problem. However,
if we intend to keep using site-local protocols, which we probably do,
we should come up with a better way to do it, and leave the SE binding
to the second case. Even there the semantics aren't very well defined,
e.g. if you specify multiple input files the broker only requires one of
them to be on a close SE (at least that used to be the case, I haven't
checked lately).
There is also the technical point that for historical reasons the
replica manager code used the access point in the CESEbind to construct
the SE pathname for classic SEs, with the result that a classic SE had
to be close to some CE. That is now fixed in the new glue schema, but I
don't know if the replica management tools have been updated yet.
Stephen
|