Child pages
  • Iterative L3 Trigger Reconstruction

Versions Compared

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

Table of Contents

...

stylesquare

Objectives and timeline:

Timeline → TSG workshop at beginning of November  (MUO POG)  
  1. Need to have the new L3 reconstruction sequence implemented, following the previous work by Benjamin. 
    Status
    colourGreen
    titleDONE
  2. Once is implemented, it is needed to compare it with the current L3 sequence.  Minimal set of the studies for old/new L3 comparisons:  
    Status
    colour

...

  1. Green
    title

...

  1. DONE

    1. Efficiency versus eta, phi, pT, PU
      Status
      colour

...

    1. Green
      title

...

    1. DONE

...

    1.   + vs dxy and dz in case of displaced

...

    1. muons  
      Status
      colour

...

    1. Green
      title

...

    1. DONE
    2. Timing (using Neutrino gun or HLT physics data, depending of what is recommended by STEAM now )

...

Status
colourYellow
titleON-GOING

...

    1.  
      Status
      colour

...

    1. Green
      title

...

    1. DONE
    2. Track quality variable comparisons (Chi2, dxy, …) for DY and background like muons (QCDMuEnriched, HLTPhysics or ZeroBias if enough stat… )

...

    1.   
      Status
      colour

...

    1. Green
      title

...

    1. DONE
  1. Comparison of the new L3 tracking with the one used in the Tk muon reco  
    Status
    colour

...

  1. Green
    title

...

  1. DOne
Timeline →

...

first month of

...

2017  (TSG) 

...

 
  1. Optimize the new L3 reconstruction:

...

  1.   
    Status
    colour

...

  1. Green
    title

...

  1. DONE
    1. IO algorithm – ROI    
      Status
      colourGreen
      titleDONE
    2. OI

...

    1. algorithm   
      Status
      colour

...

    1. Green
      title

...

    1. DONE
  1. More complete set of studies on performance comparisons using MuXX OR TkMuXX as a FOM, different signal samples should be tested in order to have all possible phase spaces covered: DY, ttbar, close-by-muon , displaced muons, low pT muon (from JPsi)    
    Status
    colourRed
    titlePENDING

    1. Efficiency versus eta, pT, PU,  (for the early studies using DY and displaced muons) + vs dxy and dz in case of displaced muons   
      Status
      colourRed
      titlePENDING
    2. Timing (using Neutrino gun or HLT physics data, depending of what is recommended by STEAM now )   
      Status
      colourRed
      titlePENDING
    3. Rate comparisons for tracks passing a given pT threshold (QCDMuEnriched, HLTPhysics or ZeroBias if enough stat… )  
      Status
      colourRed
      titlePENDING
    4. Track quality variable comparisons (Chi2, dxy, …) for DY and background like muons (QCDMuEnriched, HLTPhysics or ZeroBias if enough stat… )  
      Status
      colourRed
      titlePENDING
  2. Revise/design a smarter the combination of the out-in and in-out sequence.

...

  1.   
    Status
    colour

...

  1. Yellow
    title

...

  1. ON-GOING
  2. Ideas/prelim studies on the combination between new L3 and Tk muon reconstruction   
    Status
    colourRed
    titlePENDING
  3. Have already in place in CMSSW the new L3 sequence using current geometry.

...

  1.   
    Status
    colour

...

  1. Yellow
    title

...

  1. ON-GOING
    1. Including the new module which runs the TkMu path if the New L3 does not reconstruct the muon successfully.    
      Status
      colourRed
      titlePENDING
    2. List of unused classes that could be removed from CMSSW when switching to the new L3.  

...

    1.  
      Status
      colour

...

    1. Yellow
      title

...

    1. ON-GOING

 

Final deadline for having integrated in CMSSW, March 2017.  

...

 

 

  1. Switch to new geometry and test everything still works.   

    Status
    colourRed
    titlePENDING

  2. Clean-up CMSSW by removing unused classes.    
    Status
    colourRed
    titlePENDING

Task list for the next milestone:  

...

 

  •  

...

  •  efficiencies
  •  timing
  •  track quality measurements

...

Design:

1)   Seeding: One inside-out seeding algorithm using

For now we will use 1 seed collection. When we need to know if the seed was IO or OI we can just ask for its direction. We need to think about how the ROI will be created for the seed finding.

The OI algorithm will be a combination of state based and hit based. For OI state based the errors are rescaled, including alignment. For OI hit based we need to consider how many layers should be used, especially in the overlap region.

For the IO hit based algorithm we will reuse all the current tracking code. We need to consider how the window size will be created and whether we use the error matrix when calculating the size.

2)   Regional (track) reconstruction

3)   Matching

4)   Final fit and parameter assignment

5)   Filter

 

In New Design 1 for the production part there are 2 modules (EDProducers). The first produces a collection of L3 Tracks given a L2 muon track collection. The second produces L3 Muon track links collection are matching a L3 track with the L2 track. The following diagram describes the process:

 

 

Image Removed

 

Compared to the current design, described below there are fewer calls to read/write from the event bus (described by the blue bar in these diagrams). One of the motivations to have this in 2 modules instead of one is to perform the check on whether a seed generation algorithm needs to be activated or not by checking the currently available L3 tracks.

Running the New Code

 

IO design and testing

 

  • Properly define efficiency
    Status
    colourYellow
    titleON-GOING
    •  removing all cuts from the filter to avoid including the filter efficiency on the calculation.  
    •  Check pattern recognition, by shared hits (probably not on every sample, just a cross-check of goodness)  
      Status
      colourYellow
      titleON-GOING
  •  Performance comparisons for the between Iterative L3 IO/OI and IOH/OIH/OIS:   
    Status
    colourYellow
    titleON-GOING
    •  efficiency vs eta, pT, PU, dxy, dz, chi2...  
      Status
      colourYellow
      titleON-GOING
  •  Put a combination in place: 
    Status
    colourYellow
    titleON-GOING
    •  Understanding current under-performance: revise sequence. 
      Status
      colourYellow
      titleON-GOING
  •  Performance comparisons for the IterL3 vs Cascade
    Status
    colourRed
    titlePENDING
    •  efficiency vs eta, pT, PU, dxy, dz, chi2... 
      Status
      colourRed
      titlePENDING
    •  timing measurements. (follow TSG recipe) 
      Status
      colourRed
      titlePENDING
    •  rate comparisons/ (follow TSG recipe) 
      Status
      colourRed
      titlePENDING
  •  Prepare a complete talk containing the full package ready to deliver 
    Status
    colourRed
    titlePENDING
  •  Performance comparisons for the IterL3 vs Mu50 OR TkMu 
    Status
    colourRed
    titlePENDING

    •  efficiency vs eta, pT, PU, dxy, dz, chi2... 
    •  timing measurements. (follow TSG recipe) 
    •  rate comparisons/ (follow TSG recipe) 
  •  Based on that, decide whether a combination procedure is needed or if we can drop TkMu50 once for all. 
    Status
    colourRed
    titlePENDING
    •  repeat performance comparisons.

 


  •  Switch to new geometry 
    Status
    colourRed
    titlePENDING
  •  Clean-up CMSSW by removing unused classes 
    Status
    colourRed
    titlePENDING
    •  List of unused classes that could be removed from CMSSW when switching to the new L3 

 

Design:

A two-sequence algorithm based on iterative tracking is being implemented:   (write here more details...) 

OI design and testing

The outside-in algorithm consist of four steps:

  1. Seed generation:   TSGforOI.cc
  2. TrajectoryCandidate:   CkfTrackCandidateMaker.cc
  3. TrackProducer:  TrackProducer.cc
  4. L3MuonProducer:  L3MuonProducer.cc

The important step to optimise/study is the seed generation, current implementation allows for a great number of combinations varying the number of seeds, the rescaling factors and a number of parameters, this is the default configuration as of 15/11/2016:  

UseHitlessSeeds = cms.bool(True), 
UseHybridSeeds = cms.bool(False),
adjustErrorsDyanmicallyForHitless = cms.bool(False),
adjustErrorsDyanmicallyForHits = cms.bool(False),
estimator = cms.string('hltESPChi2MeasurementEstimator100'),
fixedErrorRescaleFactorForHitless = cms.double(5.0),
fixedErrorRescaleFactorForHits = cms.double(2.0),
pT1 = cms.double(13.0),
pT2 = cms.double(30.0),
pT3 = cms.double(70.0),
eta1 = cms.double(1.0),
eta2 = cms.double(1.4),
SF1 = cms.double(3.0),
SF2 = cms.double(4.0),
SF3 = cms.double(5.0),
SF4 = cms.double(7.0),
SF5 = cms.double(10.0),
hitsToTry = cms.int32(5),
layersToTry = cms.int32(3),
maxEtaForTOB = cms.double(1.2),
maxSeeds = cms.uint32(15),
minEtaForTEC = cms.double(0.8),

First thing checked is the distance between the generated seed compared to the reconstructed hit: 

Image Added Image Added

Several parameters can be tuned in the seed generation step:

  • run hybrid / hit seeds
  • number of seeds
  • range of TOB, TEC
  • hits/layers to try
  • adjust errors dynamically or not
  • rescaling factors

Optimization 

Here is a list of checks to optimise the algorithm:

  •  number of seeds
  •  range of the TOB, TEC
  •  once on TEC, use the stereo layers only instead of a random hit? 
  •  hit/layers to try 
  •  adjust errors dynamically
  •  rescaling factors.

A number of checks have been performed to optimised the outside-in algorithm, after some basic tunes, we gain 3% w.r.t. the performance of OIS and OIH. All the plots here: https://folguera.web.cern.ch/folguera/private/MUO/161115/OI_Optimisation/

We check that increasing the number of seeds improves the efficiency, it somehow saturates at ~10 seeds (1 TSOS + 9 hit)  

Image AddedImage Added

Then we checked the overlap region between the TOB and the TEC.  In this region, after looping over TOB the maximum number of seeds is re-initialised to 0 and the count starts in the TEC, we narrowed the region with little impact in the efficiency (0.3%), this narrowing will be faster but we still need absolute numbers for that: 

Image Added          Image Added   

 

We tried using what was called Hybrid seeds (now HitlessAndHitSeeds), which is deciding whether to add a hit seed depending on the parameters of the L2:  3% gain w.r.t using only the TSOS and 5% w.r.t old L3.

Image Added

We'll now focus on the HitlessAndHitSeeds combination.  Now we will try to find an optimal choice of (dynamical) rescaling errors. 

 

Lastly we check if we can make use of stereo layer hits instead of mono-hits in the TEC.  

 

Performance results:  

IO design and testing

Taking all the parameters from the tracker community. Already in the current HLT menu, as it is used for the HLT_TkMu50 path: 

fragment.HLTHighPt50TrackerMuonSequence = cms.Sequence( fragment.HLTDoLocalPixelSequence + fragment.HLTDoLocalStripSequence + fragment.hltL1MuonsPt15 + fragment.HLTIterativeTrackingHighPtTkMu + fragment.hltHighPtTkMu50TkFilt + fragment.hltHighPtTkMuons + fragment.hltHighPtTkMuonCands )
fragment.HLTIterativeTrackingHighPtTkMuIteration0 = cms.Sequence( fragment.hltPixelLayerTriplets + fragment.hltIter0HighPtTkMuPixelTracks + fragment.hltIter0HighPtTkMuPixelSeedsFromPixelTracks + fragment.hltIter0HighPtTkMuCkfTrackCandidates + fragment.hltIter0HighPtTkMuCtfWithMaterialTracks + fragment.hltIter0HighPtTkMuTrackSelectionHighPurity )
fragment.HLTIterativeTrackingHighPtTkMuIteration2 = cms.Sequence( fragment.hltIter2HighPtTkMuClustersRefRemoval + fragment.hltIter2HighPtTkMuMaskedMeasurementTrackerEvent + fragment.hltIter2HighPtTkMuPixelLayerPairs + fragment.hltIter2HighPtTkMuPixelSeeds + fragment.hltIter2HighPtTkMuCkfTrackCandidates + fragment.hltIter2HighPtTkMuCtfWithMaterialTracks + fragment.hltIter2HighPtTkMuTrackSelectionHighPurity )
fragment.HLTIterativeTrackingHighPtTkMu = cms.Sequence( fragment.HLTIterativeTrackingHighPtTkMuIteration0 + fragment.HLTIterativeTrackingHighPtTkMuIteration2 + fragment.hltIter2HighPtTkMuMerged )

Optimization 

Things to be checked: 

  •  Test extreme cases to make sure things work Unknown User (sfolguer):  large/tiny region using TTBar or WJets
  •  Produce a ROC curve efficiency vs timing
  •  Define the working point and close this topic.

 

Need to optimise the definition of the ROI: the current choice seems to match no other choice in the HLT, check if it is the best choice:    
MasterMuonTrackingRegionBuilder = cms.PSet(
Rescale_eta = cms.double( 3.0 ),
Rescale_phi = cms.double( 3.0 ),
Rescale_Dz = cms.double( 4.0 ), #Normally 4
EscapePt = cms.double( 3.0 ), #Normally 1.5 but it should be at least 8 for us
EtaR_UpperLimit_Par1 = cms.double( 0.25 ), #Normally 0.25
EtaR_UpperLimit_Par2 = cms.double( 0.15 ), #Normally 0.15
PhiR_UpperLimit_Par1 = cms.double( 0.6 ), #Normally 0.6
PhiR_UpperLimit_Par2 = cms.double( 0.2 ), #Normally 0.2
UseVertex = cms.bool( False ), #Normally False
Pt_fixed = cms.bool( False ), #Normally True
Z_fixed = cms.bool( False ), #True for IOH
Phi_fixed = cms.bool( True ), #False for IOH
Eta_fixed = cms.bool( True ), #False for IOH
Pt_min = cms.double( 3.0 ), #Is 0.9 for Tau; normally 8 here
Phi_min = cms.double( 0.1 ),
Eta_min = cms.double( 0.1 ),
DeltaZ = cms.double( 24.2 ), #default for tau: 24.2, for old IOH: 15.9
DeltaR = cms.double( 0.025 ), #This changes for different iterations. for old IOH: ?
DeltaEta = cms.double( 0.04 ), #default 0.15
DeltaPhi = cms.double( 0.15 ), #default 0.2
maxRegions = cms.int32( 2 ),
precise = cms.bool( True ),
OnDemand = cms.int32( -1 ),
MeasurementTrackerName = cms.InputTag( "hltESPMeasurementTracker" ),
beamSpot = cms.InputTag( "hltOnlineBeamSpot" ),
vertexCollection = cms.InputTag( "pixelVertices" ), #Warning: I am not generating colleciton. Vertex is off anyway
input = cms.InputTag( 'hltL2Muons','UpdatedAtVtx' )
)
fragment.HLTPSetMuonTrackingRegionBuilder8356 = cms.PSet( 
  Rescale_eta = cms.double( 3.0 ),
  Rescale_phi = cms.double( 3.0 ),
  Rescale_Dz = cms.double( 3.0 ),
  EtaR_UpperLimit_Par1 = cms.double( 0.25 ),
  EtaR_UpperLimit_Par2 = cms.double( 0.15 ),
  PhiR_UpperLimit_Par1 = cms.double( 0.6 ),
  PhiR_UpperLimit_Par2 = cms.double( 0.2 ),
  UseVertex = cms.bool( False ),
  Pt_fixed = cms.bool( False ),
  Z_fixed = cms.bool( True ),
  Phi_fixed = cms.bool( False ),
  Eta_fixed = cms.bool( False ),
  Pt_min = cms.double( 1.5 ),
  Phi_min = cms.double( 0.1 ),
  Eta_min = cms.double( 0.1 ),
  DeltaZ = cms.double( 15.9 ),
  DeltaR = cms.double( 0.2 ),
  DeltaEta = cms.double( 0.2 ),
  DeltaPhi = cms.double( 0.2 ),
  maxRegions = cms.int32( 2 ),
  precise = cms.bool( True ),
  OnDemand = cms.int32( -1 ),
  MeasurementTrackerName = cms.InputTag( "hltESPMeasurementTracker" ),
  beamSpot = cms.InputTag( "hltOnlineBeamSpot" ),
  vertexCollection = cms.InputTag( "pixelVertices" ),
  input = cms.InputTag( 'hltL2Muons','UpdatedAtVtx' )
)
Making a ROC curve:

Efficiency (WJets) vs Timing (HLTPhysics)

## Eficiency:
ssh -X lxplus
/afs/cern.ch/user/f/folguera/workdir/MUO/CMSSW_8_0_19/src/MuonHLT/MuonHLTDebugger/test/ROI
cmsRun myHLT_Mu50_withL3_IO_0p6_0p6_0p6_Fix.py &>myHLT_Mu50_withL3_IO_0p6_0p6_0p6_Fix.log&
## Timing:
ssh -X vocms003
cd /data/user/folguera/CMSSW_8_0_19/src/TimingScripts/test/ROI
nohup taskset -c 1 cmsRun myHLT_Mu50_withL3_IO_0p2_0p2_0p2_Fix.py >& myHLT_Mu50_withL3_IO_0p2_0p2_0p2_Fix.log&

 

Image Modified

16/10/10 22:49 :  picked dynamic 0p2 size, increasing pt_min to 3.0 GeV and improved timing, check if I can improve offs. 

Performance studies

All the plots can be found here:  https://folguera.web.cern.ch/folguera/private/MUO/

...

WJets

Image RemovedImage Removed

Image RemovedImage Removed

DYJets
Image RemovedImage RemovedImage Removed
JPsi
Image RemovedImage RemovedImage Removed
TTJets
Image RemovedImage RemovedImage Removed
Timing 

Image Removed

 

Combination procedure

Other issues and tests

efficiency drop for phi ~ 2 rads.

A drop in efficiency around 2 radians have been observed in both algorithms:

Image Added

We