simulation in babar - desysimulation in babar concezio bozzi infn ferrara desy/ecfa workshop cern,...
TRANSCRIPT
Simulation in Babar
Concezio BozziINFN Ferrara
Desy/ECFA workshopCern, 15/11/2001
http://www.fe.infn.it/~bozzi/G4cern.ppt (or .pdf, .ps)
Outline
• Reminder of Babar computing
• Geant4 and Babar simulation
• Simulation production management
• References: – Babar reports at G4 workshops and reviews (D. Wright)
– Babar reports at CHEP01 conference
– Babar Simulation NIM paper (in preparation…)
Physics Framework
• Data taking since Oct 1999
• Luminosity greater than design (>4 1033 cm-2 s-1) regularly achieved
• Yearly integrated luminosity expected to grow in the future:– 25fb-1 in 2000– 40fb-1 in 2001– 225fb-1 in 2005
620fb-1 in total
Data storage• Objectivity OO Database technology• Hierarchical structure
• Disk & tape (HPSS) storage
RAW (50kB/evt)
REC (150kB/evt)
ESD (5kB/evt)
AOD (5kB/evt)
TAG (0.5kB/evt)
Navigation DBs
Data 1999+2000: 128TB (7.3 on disk)~27500 files
MC 1999+2000: 75TB (3.1 on disk)~25300 files
NOW: 350TB
Mic
roMin
i
The Babar computing model
OBJY main event store• Filtered events (“skims”)
– Pointers to real events– Little overhead– Only navigation DBs needed
• Powerful tool for hosting a large fraction of data at SLAC
• Exportation of skimmed collection very inefficient– OBJY limitation of
64kDB/federation– Large DBs (2-10GB)
– 1 collection: several hundreds of GBs data
• Introduction of a compact format based on ROOT (KanGA)– Self-contained events– Significant overhead
• In parallel:– Develop multiple federations– Smaller DBs (0.5 GB)– Self-contained physics streams
(20) for data distribution at remote sites
• Drop Kanga if new system OK
The Babar computing model (2)
• Multi-Tiers (a-la-LHC)• Tier A: main data
repository and distribution center, all data available, all kinds of analyses possible (SLAC, in2p3, ral, infn)
• Tier B: regional, data for physics analysis (micro)
• Tier C: institutes, universities, small sample of data
3121991318254Band Mb/s
38624616210166Sum
704429169Raw
232148985422Rec
4730191820Mini
3321141115Micro
4.7321.11.5Tag
20052004200320022001(TB)Expected data to be xferred to TierA
Transfer of MC: multiply x2-x3
Sketch of data distribution
Geant4 simulationGeant4 embedded within BaBar software• BaBar Framework - outermost layer
– Handles input/output from database, files, screen, etc.– Processes data by creating objects of classes containing
BaBar code (modules) which • take data from a framework container• run specific algorithms• put results back into container for use by subsequent modules
• BOGUS - next layer down– Performs G4 based simulation within framework– Passes results to detector response module then to
reconstruction module
• Geant4 toolkit - next layer down– used to do large part (but not all) of simulation
G4 in the Babar simulation
B - BaBar
O - Object-oriented
G - Geant4-based
U - Unified
S - Simulation
G4 features used in Babar• Geometry
– Constructed Solid Geometry (CSG) solids– boolean volumes (using CSG constituents)– G4 materials
• Hit scoring– G4 virtual sensitive detector– G4 sensitive detector manager
• Physics processes– EM physics (E < 10 GeV)– G4 hadronic processes also used (E < 10 GeV) but not
fully tuned yet
• Visualization (openGL)
Modeling of physics processes• Particle range cut (instead of energy threshold)
– 0.22mm (δ rays produced in DIRC above Cherenkov threshold)– Eventually allow for material dependence?
• EM physics processes:– Photons: G4ComptonScattering, G4GammaConversion,
G4PhotoElectricEffect– Electrons/positrons: G4eIonisation, G4eBremsstrahlung,
G4eplusAnnihilation– Muons: G4MuIonization, G4MuBremsstrahlung, G4MuPairProduction
– Hadrons: G4Ionisztion– All charged particles: G4MultipleScattering
• Hadronic physics processes (LE models):– Elastic:G4HadronElasticProcess (G4LElastic)– Inelastic, eg p+: G4PionPlusInelastic Process (G4PionPlusLEInelastic)
G4 features NOT used in Babar• Tracking: BgsTransportation
– Developed at SLAC because of required precision, fast execution (reduced number of calls to B field), no bias in tracks
– no Runge-Kutta integration, uses only straight lines and helices and is fast and accurate for nearly uniform BaBar field
– 4 criteria for minimum step
– Intersection with G4 volumes by iterative procedure
– other 9 G4 steppers being tested (RKG3_stepper appropriate?)
• Detector response (most of BaBar detector response code written before G4 was available)
• Persistence (BaBar database and framework use their own persistent objects)
Geometry
Implementation of G4 specifics solids:(polyhedra, polycone, hype, elliptical tube)
PEP geometry and SVT
Geometry diagnostic tools
• Material scanner– returns radiation length, interaction length, path
length, theta, phi, coordinates of track exit point as an n-tuple
• Geometry checker– specifies straight lines through detector
– returns boundaries along line
– useful for finding volume overlaps
Material validation• Bremsstrahlung in
Bhabha events• Conversions in gamma-
gamma events
Tracking performance: µ+µ-, bb
D0 mass
D0 lineshape fits
Sample Mean (GeV) RMS width (GeV) Chi2/dof2001 Data 1.86351 ± 0.00007 0.00697 ± 0.00019 32.1/32
SP4 Geant4 1.86406 ± 0.00006 0.00684 ± 0.00009 27.8/25SP4 Geant3 1.86514 ± 0.00005 0.00476 ± 0.00014 39.5/24
SP3 1.86504 ± 0.00005 0.00491 ± 0.00016 29.5/28
EMC comparisons• G3/G4 show no significant differences
IFRvariables
µ+µ-
Summary of comparisons
• Detector Material Model validated for tracking region of detector
• Resolution and decays– reconstructed mass resolution looks good – track resolution needs more work
• EMC performance similar in G3/G4– G4/data ongoing
• IFR: more tuning needed– not worse than G3
Run-time performance• Fast
– except for initialization time, as fast as G3– initialization time = 30 cpu seconds
CPU times using SLAC Linux farm (seconds/event) : Event type Bogus (G4) Bbsim (G3)B0B0bars 5.43 6.24mu pairs 1.10 0.62Bhabhas 8.52 7.58
• Robust– low crash rate (~1ppm)– No significant memory leaks– running on three platforms (Linux, Sun, Dec)
• Production using G4 has begun– >100M events produced so far– >50% produced outside SLAC
Simulation Production Management• 3-staged simulation production system:
– Simulation using Geant4 (or 3)– Mixing: add detector response and background signals– Reconstruction: like in real data
• Events output from one stage are used as the input to the next stage
• Spread production on several sites• Many user requests, many production jobs
(100,000), many events (500M)– 1 CPU for about 150 years -> use 500 CPUs
• Most production done on >800MHz Linux (bi-processor) machines, but also on Sun and OSF
Structure of the system
• Collect and centrally manage requests for simulation– Web interface to Oracle DB
• Distribute requests to production sites– Separate request into runs (each of them requires 8 hours of
computing for all 3 stages), usually 2K to 10K events– Assign to production sites
• Manage jobs at each site– Set of perl scripts, interacting with Oracle DB (proxy)
• Import events to SLAC– Automatic procedure, perl script– Closed DBs extracted from OBJY and transferred to SLAC– Automatically imported into the SLAC federation
Status of G4 MC production• Start: Aug. 15th (4 sites)
• 15 sites in production
• More to come
• Nominal rate: 3Mevts/day– 2M outside SLAC
Conclusions
• Geant4 toolkit embedded in Babar simulation– Conservatively for now…– Fast and robust
• Better or at least as good as Geant3– Material scan– Tracking– EMC– IFR: still room for improvement, tuning…
• Huge simulation production efforts (>80Mevts/month) started in summer 2001