L3 Filtering: status and plans for the near future D Collaboration Meeting: 25 th April 2002 Dan...
-
Upload
miles-cobb -
Category
Documents
-
view
217 -
download
0
description
Transcript of L3 Filtering: status and plans for the near future D Collaboration Meeting: 25 th April 2002 Dan...
L3 Filtering: status and plans for the near future
D Collaboration Meeting: 25th April 2002
Dan Claes and Terry Wyatt,on behalf of the L3 Algorithms group.
For more details see, e.g., L3 talk at trigger workshop:http://www-d0.fnal.gov/computing/algorithms/level/docs/L3_workshop.ppt
Within a month:
L1 L2 L3 tape 500 Hz 30-40Hz
Currently:L1 L3 tape . 100 Hz 20-30 Hz
• Factor ~ 5 rejection needed
• Calorimeter-based filtering only (jets, electrons, taus)
• Next steps (p11.06 release) to commission:
Muons, central tracking, track-based primary vertexing, calorimeter non-linearity and hot cell killing.
Time budget for L3 i/o, event building, filtering 100 nodes/500 s = 0.2 s
Level 3 Jargon
• Tool: Does the real work– Unpacks raw data, finds tracks, clusters – Identifies physics objects (e, , , jet ,, W, Z)
• Filter: Applies (simple) cuts on objects • e.g. pt() > 5 GeV
• Filter Script: Defines a L3 trigger condition– Logical .AND. of one or more filters
• e.g. pt() > 5 GeV .AND. pt(jet) > 10 GeV
– If all filters in a script are .TRUE. trigger is satisfied and event is recorded
• ‘Mark and Pass’: – Selects unbiased sample of input events to be recorded
In order to save processing time:• Only a partial reconstruction of each event is performed, depending on the
L1/L2 trigger information• For each L1/L2 trigger that fires:
– One or more L3 filter scripts run– Each script calls the filters/tools necessary for the trigger decision
• Tool results are kept in case they are needed again in that event
Details of which filters/tools are called by each script:
• determined by the triggerlist
• and performed by: ScriptRunner Author: Moacyr Souza (Fermilab/LAFEX)
L3 central code management: Jon Hays (Imperial)
Short term goal to run a filter L3FMuon• with ‘loose’ quality• cutting on pT
Mainly needed for single muon triggers(which currently have a L1 prescale of ~40)
L3TMuon (local muon track reconstruction)
original author: Paul Balm (NIKHEF) current responsibles: Christophe Clement (Stockholm) , Martijn Mulders (Fermilab), Martin Wegner (Aachen)
+ L3TMuon uses much of the ‘offline’ muon reco code
L3TMuon
• Dynamic unpacker commissioned (p11.04.00)• Longstanding memory leak fixed (p11.04.00)• Keeping up with updates to the ‘offline’ muon reco
code:– new memory leak and timing problems– temporary fixes by adjusting rcp parameters
L3FMuon has been run online in special runs
Timing
• p11.04.01 default rcp parameters (maxopt version running on 1 GHz PIII)• mean time/event ~200 ms• when run online ~50 out of 200,000 events took more than 30 seconds to process
• p11.04.01 parameters tuned to reduce time taken: - number of A/BC segments considered - extrapolation step size - number of track-fit iterations • mean time/event ~20 ms• eliminates time-outs
Efficiency (pT = 5 GeV Monte Carlo single muons)
‘Loose’ L3 muon
‘Tight’ L3 muon
Notes:‘Loose’ efficiency ~80% (cf. geometrical acceptance ~90%)Poor ‘tight’ efficiency in central region (track fit fails to converge – also seen in ‘offline’ reco.)
Default RCP parameters
‘Tuned’ RCP parameters
Rejection measured on data: single muon test runs
defaultparameters
tuned parameters
central forward
• Thursday 24th April: – another single muon test run taken (with tuned parameters)– showed no signs of timing or memory problems
• Get L3FMuon running online full-time– Global_CalMuon6.00 exits with loose L3FMuon
hanging off L1 single muon and muon-jet triggers (100% Mark&Pass)
Next Steps for L3FMuon
Next Steps for L3FMuon• Fix muon reco memory leak • Optimise parameters for L3:
– Memory/timing efficiency/rejection– Find out why some events take so long to reconstruct
• Stricter procedures for production releases of ‘offline’ muon reco software– Including a specific requirement for L3FMuon tests
BEFORE code changes released to production branch• Comparison L2 L3 (e.g., efficiency, momentum
resolution)• Longer term: We need a serious analysis of the
cost/benefit of retaining/breaking the link between L3TMuon and muon reco
Recent progress in L3 central tracking
• Offline quality unpacking and geometry for L3• Improved SMT-CFT matching
– stand-alone tracking filter– track-based primary vertex tool
SMT unpacker:• Fully dynamic, parameterized pedestals, noisy strip killing
CFT unpacker: • replace global threshold with individual channel thresholds • up to date thresholds and cable maps
Offline quality geometry implemented for SMT and CFT in L3
(Released in p11.06.00)
When improved thresholds/cable maps/geometry are available:• requires no code changes• but care in archiving/version-tagging (general problem!)
SMT/CFT Unpacking and Geometry
Principal author: Robert Illingworth (Imperial College)
Level 3 Global Track Finding author: Daniel Whiteson (Berkeley)
• Find axial CFT tracks• Match stereo CFT clusters• Extend into SMT• Require 8/8 axial CFT hits if no matched axial SMT hits• Require 7/8 axial CFT hits if 3 matched axial SMT hits• If CFT axial/stereo match fails:
• CFT-SMT match done in xy-only• but SMT stereo information still used to give 3D tracking
(new feature implemented in p11.06.00)
CFT Tracking Algorithm - L3TCFTTracker
Principal author: Ray Beuselink (IMPERIAL) P11 tool certification: Robert Illingworth and Chris Barnes (IMPERIAL)
Current L3 global tracking performance
track (rad) z at dca (cm)
Number of axial hits on track
CFTonly
Number of stereo hits on track
Comparison of L3 and offline tracking
number of tracks track (rad)
q/pT (GeV-1) dca (cm)
Timing for global track tool
Time per event (ms) (d0mino - debug version)
+ about 8 ms per event for unpack tools (1 GHz P3 – maxopt)
Next steps for central tracking
• Test p11.06.00 version on most recent available data
• Get it running online full-time!• Measure efficiency vs. rejection rate of
stand-alone track filter on, e.g., single muon events
A possible plan for filtering on single muon triggers
• EITHER: Loose L3 muon • OR: Central track
– (i.e., using ‘redundancy’ to improve efficiency)– N.B. Tracking will not be perfect for a long time– (If you don’t like this, you can always exclude these event by using the
L3 trigger names)
• Longer term:– May need to require track-muon match (at least for low pT)
Tool exists (Paul Balm)
– Also investigate track-Calorimeter MIP match Tool under development (Martin Grünewald)
We could require an .OR. of:
L3 Primary Vertex needed soon!
author: Guilherme Lima (UERJ/Brazil)
Has yet to be fully tested on REAL DATAOpportunity for new person to get involved!
1) Histogram technique using SMT hits
2) Track-based L3TVertexFinderauthor: Ray Beuselinck (Imperial)testing: Chris Barnes, Per Jonsson (Imperial)
Recently upgraded to use either CFT or Global candidate tracks as input
N.B. Marseille group (Arnaud Duperrin, Mossadek Talby,Eric Kajfasz) hope to be actively involved in testing tracking, vertexing.
Monte Carlo Performance
efficiency purityZ 75% 99%tt 95% 95%
(to find a vertex within 1 cm of the correct primary)
z residual in Z events
z (cm)
Reconstructed vertex position in real data
z (cm)
~40 % of eventshave a vertexreconstructed
Next steps:- tune cuts on numbers of hits required (especially for SMT stereo hits)- finish certification
L3TEle Electron Tool authors: Volker Buescher (Mainz) Ulla Blumenschein (Mainz)
Current filter:• simple 0.25 cone• applying cuts on
• ET
• e.m. fraction (>0.9)• transverse shower shape
, : energy weighted cluster axis position
i
iii
EE
r22 )()(
Efficiency in Monte Carlo Events
Electron ET (GeV)
Electron ET (GeV)
Comparison of data and single electron Monte Carlo (trigsim):
trigger turn-on curve
Electron ET (GeV)
Efficiency (%)
Single electron triggers
• At L1 we have two (unprescaled) single electron triggers:– CEM(1,10)– CEM(2,5)
• At L3 we run the same two single electron filters:– pT > 15 GeV, emfrac > 0.9
– pT > 12 GeV, emfrac > 0.9, shower shape– Combined rejection factor ~3.5
• We do a similar thing with CEM(1,5) (heavily prescaled)
Further developments for single electron triggers
• More sophisticated treatment of transverse and longitudinal shower shape– Studies in progress
• Add in parallel to the two current filters: – Higher pT cut and softer e.m. fraction cut?– Stand-alone track filter?– Matched track + looser e.m. cuts?– (Matched pre-shower + looser e.m. cuts)?– Do we have enough L3 trigger bits (256)? – Alternative: have one filter that combines all available information
(with details of the trigger decision stored in L3PhysicsResults)• When we have to cut harder (soon) this redundancy will
help to maintain:– High efficiency– Small systematic error
L3TJet Tool author: Volker Buescher (Mainz) high precision calorimeter readout available at L3
sharpen the turn-on curve
• running online stably since early Sept’01• rejection factors 20-50 for different triggers
e.g., fraction of events passing L3FJet(1,15)
pT of leading offline JCCA jet (GeV)
Next Steps (p11.07)
• Use primary vertex ?
• Calorimeter non-linearity corrections implemented in calorimeter unpacker (Marumi Kado)
• Killing of hot cells implemented (Marumi Kado, Gregorio Bernardi have implemented NADA into L3)
Running online since Jan-2002Example:mu1ptxatxx_CJT(1,3) + L3Tau (pT > 10 GeV)gives rejection factor ~ 5.5
Next step: switch on tracking
Z QCD
L3FTauHadronic Level3 TauTool author: Gustaaf Brooijmans (Fermilab) current responsible: Yann Coadou (Upsala)
Based on calorimeter jet shape variables
Other tools on a longer timescale• missing ET
• tools to associate objects in different detectors (e.g. track to muon)
• cps and fps cluster finding and unpacking• b-tag: impact parameter, secondary vertex• tools to calculate "physics" quantities
– (e.g., inv. mass, delta_eta)• tools to identify physics event types
– (e.g., W, Z, stream definitions)– Keep raw data on reco output of W/Z candidates?
Many opportunities for new people to get involved!
L3 Monitoring• L3 filter statistics for each trigger available to
shift crew via daq_monitor• L3 reconstruction results written out with the
raw data• l3fanalyze program: produces rootuple
– Used offline for:• testing new code versions• checks of data quality
– Plan to run online as ‘examine’ – use root to fill monitoring histograms from rootuple
Monitoring
• Extra person(s) urgently needed to work on this! – macros to define monitoring histograms– common job submission on standard test samples to
exercise all the L3 tools/filters (and L2?)– migration to online – "bit-wise" on/offline check – matching L3 objects to MC/L1/L2/reco objects
• L3 monitoring needs to get a lot more systematic and routine!
Can we do more sophisticated online monitoring in the L3 nodes?
• For example, collect histograms, measure efficiencies– L3 does a pretty complete reconstruction of the data
• Make use of the 95% of the events that we reject? – Measure trigger turn-on curves (for L1 and L2 as well as L3)– Do background studies– (Why write out events and have the huge overhead in having to run offline reconstruction and storing them permanently if they are needed for relatively simple operations that can be performed adequately in L3?)– Write out a stream with L3PhysicsResults and no raw data?– Write out an ‘insurance’ or ‘not for reco’ stream?
• Best way to concatenate results from monitor processes running on each of the 100 L3 farm nodes not worked out yet.
Will require extra resources at L3, but the potential return (in terms of spotting trigger problems and in saving offline resources) might make this a very cost-effective investment.
This will also be the case if we find that lack of cpu power is limiting the sophistication of the event reconstruction and/or filtering that is possible in L3.
L3 central infrastructure: opportunities for new people to get involved
• Scriptrunner + central L3 code infrastructure, release management
• Calibration/alignment technical infrastructure
• L3 reconstruction results on thumbnail• Streaming• Development of "user" and "physics
analysis" tools
Conclusions, outlook.• Currently:
– Calorimeter-based filtering (jets, electrons, taus) only
– Factor ~ 5 rejection needed at L3
• Next steps (p11.06) to commission:– Muons, central tracking, track-based primary vertexing,
calorimeter non-linearity and hot cell killing.
• Within a month: – DAQ improvements L3 input rate 500 Hz.
• In the next few months:– L1 tracks, L1 calorimeter acceptance and large tiles
– L2 rejection increasing
– L3 input rate 1000 Hz
– Higher luminosity
– Lots of interesting challenges and scope for new people
To find out more:
• L3 Algorithms web-pages:http://www-d0.fnal.gov/computing/algorithms/level3/home.html
• L3 Algorithms working group meetings take place every week:
Wednesday 14:00-15:30 in the Farside
• Talk to Dan Claes ([email protected]) or Terry Wyatt ([email protected]) about the opportunities to get involved!