LHC Computer Grid - Rollout
> [mailto:[log in to unmask]] On Behalf Of Ben Waugh said:
> As someone (maybe you?) has pointed out earlier, if the ERT is to be
> useful for ranking sites then it must be calculated in the
> same way for
> each site. That being the case, would it not be more sensible for the
> algorithm to be defined by the RB or the UI?
That's really two different questions. The ERT is, at least at the
moment, one of the defined attributes in the Glue schema, hence
information providers for the Glue schema have to provide a value, and
ideally would do the best job they can given the well-known problems -
the old info provider did a reasonable job given the constraints, the
new one doesn't.
Separately there's the question of what is used to rank CEs and decide
where to submit the job. That is defined by the user, and can be any
expression involving (pretty much) any value in the schema. If the user
doesn't specify anything there is a default expression in the UI
configuration file. I've argued repeatedly over the years that that
default should *not* be the ERT because of all the problems it causes,
it's certainly possible to define expressions from other things like
FreeCPUs and WaitingJobs which are more robust, but in practice nearly
everyone seems to have stuck with ERT and then complained that it
doesn't work!
What we are now proposing is to add a VO-dependent ERT to the schema,
and Jeff Templon has code to provide an estimate which depends on the
recent history of the batch system (and hence can only be calculated at
the site) which should have a better behaviour, although we'll have to
get experience with it to see how it works.
Stephen
|