Attending: Sam, Jens (chair+mins), John B, Steve, Daniel, Alastair, John H, RobC, Winnie, Elena, Brian, Robert, MattD, Govind

0. Operational blog posts

   No new blog posts.  Of course STFC may not be allowed because of the Purdah.

1. Alastair has kindly volunteered to present some thoughts on xroot
   caching, so he will be joining us to, er, present his thoughts.

   Alastair presented his presentation.  He had talked to Wei Yang, to Andy H (- "easy, no bugs")
   and to Marcus who is still around in some sense somewhere.

   Mainly taking aim at "small" sites; size depends on number of TB with < 100 being the smallest;
   between 100 and 400 being the next, and then above - this volume does not include the groupdisk
   cap.

   Aim is also to reduce effort requirements at such sites.  Local user access can still be
   maintained independently.  Generally ATLAS services should have failover so local failure
   should just slow the job down as it recovers data via the federations.

   Different systems targeted - there is cache or proxy with disk version, a memory version, and
   a "disk proxy cluster" or somesuch.  Note that Brian Bockelman is due to give a talk on caching
   at WLCG.

   Some debate about the access patterns, e.g. CEPH needs to access whole object or access will be
   relatively inefficient.  If the federation facility needs to recover data from different
   flavours - memory cache and CEPH, say, it may just have to cope/live with the relative
   inefficiency.  ROOT also has a T3 cache facility which should perform prefetch and
   optimisations as needed, also collecting multiple accesses etc.

   Obviously it would be good to do some experimenting, and would also need monitoring to see
   how efficiently it runs.  A related question is how this would affect/benefit other experiments,
   as obviously we don't generally want to run ATLAS-only services (unless you're a small enough
   site to only support a single VO.)  Sam had been keen to test with/at Oxford so we would get
   potentially CMS experiences as well, but Kashif had not been willing/able/available.  Matt
   suggests using hepsysman to make progress on this topic.  Matt may also be able to test local
   access (even though Lancaster is not a small ATLAS site in this sense.)  JohnH is also willing
   and could do some work provided it doesn't require too much effort, but will not be attending
   hepsysman.

2. https://indico.cern.ch/event/609911

   Agenda for the data session is now pretty much complete, although there may be room to suggest
   minor items.

3. AOB

   NOB


John Hill: (31/05/2017 10:04:18)
Yes
Samuel Cadellin Skipsey: (10:05 AM)
I hear it's perfect and was given to us by the Gods themselves.
Jens Jensen: (10:07 AM)
:-)
Matt Doidge: (10:44 AM)
Perhaps we could lend a hand at HEPSYSMAN?