UKQCD NeSCAC Irving, 24/1/061 January 06 UKQCD meeting Staggered fermion project Alan Irving University of Liverpool
UKQCD NeSCAC Irving, 24/1/062 Who and why? ‘Who’ is based on initial proposal Cambridge: Ron Horgan,… Edinburgh: Alistair Hart,.. Glasgow: Christine Davies, Alex Dougall, Eduardo Follano,…. Liverpool: Eric Gregory, Alan Irving, Craig McNeil, Chris Richards,.. Oxford: Mike Teper,… Swansea: Chris Allton,… + Mike Clark (algorithms), Peter Boyle (support), Balint Joo (Chroma),… ‘Why’ : General aim is to build on existing Asqtad calculations to: Explore new areas that require new levels of computing power Push down systematics on experimentally relevant quantities Main advantages are: Access to light dynamical quark masses with useful level of statistics Large volumes with small lattice spacings Discretisation errors under ‘good control’
UKQCD NeSCAC Irving, 24/1/063 Physics goals Light hadron spectrum –Complementary data sets to MILC (incl s-fine) –Different/more operators to MILC, –Longer runs, MEM analysis,.. –Quark masses with 2-loop matching Flavour singlet physics – , ’ masses –Singlet scalar and glueball masses –Nucleon matrix elements –Chiral condensate, topological susceptibility Heavy quark physics –Upsilon fine structure (improved NRQCD) – B Physics: B l with moving NRQCD (low q 2 ) –Charm with highly improved action (see CTHD)
UKQCD NeSCAC Irving, 24/1/064 Proposed runs Production of improved staggered as ‘proposed’ last year: –m ud /m s =1/5 –0.125fm, 24 3 x64, 30K traj –0.09 fm, 32 3 x96, 20K traj –0.06 fm, 48 3 x144, 2K traj Resourcing ‘approved’ by UKQCD last year: –5 racks (50% of QCDOC?) subject to verified code Current performance: –Long, medium lattice (.006/.03) run 10 configs/rack/day (7 months on 2 racks) –MILC spreadsheet estimate was 0.30 Tflop years (4.4 months on 2 racks) –Could be improved by ~2-3 tmes with Mike Clark RHMC-algorithm improvements? –MILC may move to RHMC?? Proposed near term use of QCDOC on staggered programme: –2 racks on long medium lattice run? (7 months) ( 3 months?) –4 racks on ‘extra fine’ lattice? (4 months for 2K of MILC.0072/.018) ( 2 months?) –1 rack (as available) on measurement including singlet* operator measurements? * 64 noises, 20 3 x48, m q =.02: 30 hours on 1P4 with 2Gb
UKQCD NeSCAC Irving, 24/1/065 Status of Asqtad configuration generation Staggered project web pages: – ERED/ Initial runs on 24 3 x64, β=6.75, ma = 0.006/0.03 –‘long’→30K trajectories, medium lattice for singlet physics, autocorrelation,.. Rack 39 running ‘signed off’ (help from Peter B) –Detuned binaries (ZS) –Uptime (between power cycling) now in weeks (~ 3 weeks?) –148/138/126 Mflops per node for Solver/Gauge Force/Fermion Force. – Chris Richards driving this rack. –Reproducing OK, m PS checks with MILC data (~296 MeV) – doing 10% reproducbility runs ( ~ 2weeks) – trajectories, so 250 configs (separated by 6 traj) so far. –Mean acceptance 65% (50 steps of.01) –Int AC time 7.3 +/- 2 on plaq.
UKQCD NeSCAC Irving, 24/1/066 Other racks… MILC experience (DoE machine) 2 rack machine running 40 3 x96, /0.031 ( a=0.125fm ) –326 traj, reruns of 18%, overall utilisation 63% 4 rack machine running 48 3 x144, /0.018 ( a=0.06fm ) –402 traj, reruns of 30%-20%, overall utilisation 51% Performance –326 traj in 3500 hours on 2 racks scale: 70 mins cf 22 mins (UKQCD)?? Rack 44 not so good… –No configs so far – Reproducibilty failures – Craig working with QCDOC-support since December. – Advice has been to shake out this rack/code combination before taking on another rack?