People on Vidyo: Andrea Sciaba, Derek Weitzel, Diego Ciangottini, Dirk Duellmann, Eric Fede, Riccardo Di Maria, Xavier Espinal, David Smith, Martin Gasthuber, Oxana Smirnova, Frank Wuerthwein, Ilija Vukotic, Laurent Duflot, Nikola Hardi, Nikolai Marcel Hartmann, Stephane Jezequel, Teng LI, tigran
- Dr Teng LI (University of Edinburgh, UK) - XCache and future plans of diskless sites at GridPP:
- the resources (mostly consumed by ATLAS, CMS) are highly concentrated
- the current simplest solution: diskless site with direct access to a nearby large SE
Birmingham (attached to Manchester), Sheffield (attached to RAL ECHO), Cambridge (attached to QMUL), ECDF (none-diskless site, running XCache)
- comparison of Birmingham production queue and ECDF analysis queue: cache hit rate at 50-70 % and production queue is more ‘cache-able’
- metric collector container developed at ECDF: XCache logging, cached files and OS, host information, cache hit/miss, cached files
- BHAM: running very smoothly for 7 months
- metrics collected from BHAM: 9310 (6.6T) files written to cache; 48426 (20.8T) read requests
- even with one node, cache will not be filled; these are production jobs using tiny amount of data
- one could increase the pressure on the site for both simulation and I/O jobs by changing the load
- the cache should be pushed to the limit before summarise
- popular files are related to HC tests
- Sheffield: analysis jobs use direct I/O; production jobs use copy-to-scratch -> huge difference
- indirect access good as long as you keep connection open over (even long) time
- better to have direct access when it comes to cache
- summarising: 3 diskless sites attached to nearby large sites; Cambridge and Birmingham running XCache; positive experience with Birmingham; Sheffield under investigation; if more sites joining, plan to investigate cache federation
- cache federation should be followed up since the plan is not clearly defined
- wrt CMS: RAL should have tested with AAA