 |
 |
Hi, A quick comment on the PROOF note below. I realize that SLAC isn't going that direction, but for future reference I thought I'd mention some of what I see as a problem with the current PROOF model for university settings. At UW I've been looking into PROOF a bit recently. The current model is for a dedicated set of machines: you need to modify the inetd file on each machine that will run the PROOF executable. If your computing cluster does multiple tasks (even multiple groups) then PROOF does not play well with those other tasks. Cheers, Gordon. > -----Original Message----- > From: [log in to unmask] [mailto:owner- > [log in to unmask]] On Behalf Of Yang, Wei > Sent: Tuesday, July 03, 2007 12:27 AM > To: atlas-sccs-planning-l > Subject: summary of US ATLAS Tier2/Tier3 meeting at Bloomington (June > 2007) > > 1) Next T2 meeting at SLAC > It will be in November. We will pickup a time range and talk to > Jim Shank and others to make a final decision. > > 2) Next T2 hardware purchasing > We need to evaluate the next round of hardware purchasing ASAP, > and identify power, cooling, etc for them (and 10Gb WAN?). US ATLAS > computing management expects that T2 sites to send out purchasing order > shortly after the FY2008 funding is in place and ramp up the pledged > 2007 end-year capacity in time. > There will be a big gap between projected CPU and storage capacity > and targeted capacity from 2009. New LHC running assumptions COULD lead > to some savings in a later ramp-up of hardware. However, there are huge > uncertainties about this saving due to largely increased event size. > > 3) Integration Plan/SRM > It looks like that a SRM interface to the storage is required in > the near future. Gsiftp alone is not enough to load balance a large > numbers of concurrent requests. > There will be some US sites verification, availability and > performance tests in the middle of July. > > 4) Storage hardware model change at BNL > As storage size increase, BNL is moving away from the model that > distribute storage on CPU nodes, and has purchased 35 thumpers. > > 5) Analysis priority > People are talking about setting priorities for analysis jobs. > Several implementation methods were mentioned but none are mature. It > is especially not clear whether this should be done at site level, or > at Panda scheduling level. > > 6) DDM > CERN and all US sites have "upgraded" to DQ2 0.3. There were > initial problems everywhere, especially at CERN. People are sorting out > bugs. > At Bloomington, Wei and Alexei did an interesting test and found > that each file transfer has more than one minute of overhead. It is > recognized that there will be small file transfers. > > 7) Proof/Xrootd > Torre Wenaus gave an interesting talk about using PROOF to do > "parallel" post-Athena analysis. It prefers to access root data locally > available to the CPU nodes, though remote access also works. Xrootd > based storage work well with PROOF, though xrootd is not required. > However, since BNL (and SLAC) will not distribute storage on CPU nodes, > PROOF will have to find playground mainly in university group > computing. > > 8) Tier 3 centers and university group computing > The computing power at T1/T2s will not be enough for all US users > to do analysis. Bloomington meeting spend 1.5 days to help university > users understand ATLAS DDM and Panda systems, and to build their own > group computing or Tier 3 centers. > > -- > Wei Yang | [log in to unmask] | 650-926-3338(O)