Primary Vertex Reconstruction for Heterogeneous Architecture at CMS

(CMS DP-DP-2022/052)

This note reports on vertex reconstruction at CMS in HL-LHC conditions.

Metric definitions and more details can be found in the DP-2022/052 note (CDS record: https://cds.cern.ch/record/2839922?ln=en).

effic_vs_NumVertices.png

Primary Vertex Efficiency vs number of simulated vertices:

  • the current vertexing algorithm (black);
  • the new DA clustering in blocks and the current adaptive vertex fit (red);
  • the current DA clustering and the new weighted mean vertex estimator (green);
  • the new DA clustering in blocks and weighted mean vertex estimator (pink).

Contact: Giorgio Pizzati
effic_vs_NumVertices.png

Primary Vertex Efficiency vs $ \sum p_T^2 $:

  • the current vertexing algorithm (black);
  • the new DA clustering in blocks and the current adaptive vertex fit (red);
  • the current DA clustering and the new weighted mean vertex estimator (green);
  • the new DA clustering in blocks and weighted mean vertex estimator (pink).

Contact: Giorgio Pizzati
effic_vs_NumVertices.png

Primary Vertex Fake rate vs number of simulated vertices:

  • the current vertexing algorithm (black);
  • the new DA clustering in blocks and the current adaptive vertex fit (red);
  • the current DA clustering and the new weighted mean vertex estimator (green);
  • the new DA clustering in blocks and weighted mean vertex estimator (pink).

Contact: Giorgio Pizzati
effic_vs_NumVertices.png

Primary Vertex X coordinate resolution as a function of $ \sum p_T $:

  • the current vertexing algorithm (black);
  • the new DA clustering in blocks and the current adaptive vertex fit (red);
  • the current DA clustering and the new weighted mean vertex estimator (green);
  • the new DA clustering in blocks and weighted mean vertex estimator (pink).

Consistent trends along Y were observed

Contact: Giorgio Pizzati
effic_vs_NumVertices.png

Primary Vertex Z coordinate resolution as a function of $ \sum p_T $:

  • the current vertexing algorithm (black);
  • the new DA clustering in blocks and the current adaptive vertex fit (red);
  • the current DA clustering and the new weighted mean vertex estimator (green);
  • the new DA clustering in blocks and weighted mean vertex estimator (pink).

Contact: Giorgio Pizzati
effic_vs_NumVertices.png

Primary Vertex X coordinate pull as a function of $ \sum p_T $:

  • the current vertexing algorithm (black);
  • the new DA clustering in blocks and the current adaptive vertex fit (red);
  • the current DA clustering and the new weighted mean vertex estimator (green);
  • the new DA clustering in blocks and weighted mean vertex estimator (pink).

Consistent trends along Y were observed

Contact: Giorgio Pizzati
effic_vs_NumVertices.png

Primary Vertex Z coordinate pull as a function of $ \sum p_T $:

  • the current vertexing algorithm (black);
  • the new DA clustering in blocks and the current adaptive vertex fit (red);
  • the current DA clustering and the new weighted mean vertex estimator (green);
  • the new DA clustering in blocks and weighted mean vertex estimator (pink).

Contact: Giorgio Pizzati
effic_vs_NumVertices.png

Primary vertex (hard scattering, one per event) tagging power, defined as the probability of reconstructing the signal primary vertex and tagging it correctly as described in Section 9.4.1 of Ref [3] :

  • the current vertexing algorithm (black);
  • the new DA clustering in blocks and the current adaptive vertex fit (red);
  • the current DA clustering and the new weighted mean vertex estimator (green);
  • the new DA clustering in blocks and weighted mean vertex estimator (pink).

Contact: Giorgio Pizzati

Algorithm label

P.V. Producer Time per event [ms]

Old Vertexing

913.0 3.5

New Clustering + Old Fitter

368.0 1.3

Old Clustering + New Estimator

512.6 3.3

New Vertexing

145.7 0.7

The table reports the timings in ms evaluated on 2000 events with <PU>= 200 averaged over 10 repeated measurements for:

  • the current vertexing algorithm (black);
  • the new DA clustering in blocks and the current adaptive vertex fit (red);

  • the current DA clustering and the new weighted mean vertex fit (green);

  • the new DA clustering in blocks and weighted mean vertex fit (pink).


These measurements were run with a single process with 10 concurrent threads on Intel Skylake Gold processor, with no other processes running on the machine.

Contact: Giorgio Pizzati


Clustering Working Points

|

Overlap 0.5 Overlap 0.25 Overlap 0.15
Block Size 512 256 128 512 256 128 512 256 128 Current
PV tagging 0.94 -0.4% -0.2% -0.5% -0.2% -0.7% -1.7% -0.2% -1.8% +0.1%
Efficiency 0.77 +2.8% +7.5% -3.1% -0.5% +3.1% -4.6% -1.8% +2.7% -8.2%
Fake rate 0.13 +11% +24% -14% -4.7% +13% -18% -8.7% +10% +0.0%
Duplicate rate 0.09 +65% +177% -28% +22% +89% -34% +4.6% +79% -89%

| The table shows the different metrics (PV tagging power, efficiency, fake rate and duplicate rate) compared to the default case (overlap 0.5, block size 512) for nine different configurations with overlap fraction of 0.15, 0.25, 0.5 and block size of 128, 256 and 512.

  • Best cases for each metric are highlighted in light green when compared to the default case

  • Changing the clustering parameters the vertexing performance may tuned to better adapt to the physics case of interest. In this context, the default case was chosen because, among those with the highest PV tagging, it is the one keeping under control the duplicate rate.

  • The performance has been measured with a simulated sample of 2k events with 200 superimposed pileup (PU) interactions per event with Run4 detector conditions.

Contact: Giorgio Pizzati

|

effic_vs_NumVertices.png

Execution time of the new primary vertex producer as a function of the blocks overlap fraction for:

  • block size of 512 (black);
  • block size of 256 (red);
  • block size of 128 (green);

Contact: Giorgio Pizzati

effic_vs_NumVertices.png

Speed up factor of the new primary vertex producer, defined as the ratio between the new and the current producer execution time, as a function of the blocks overlap fraction for:

  • block size of 512 (black);
  • block size of 256 (red);
  • block size of 128 (green);

Contact: Giorgio Pizzati

-- GiorgioPizzati - 2022-10-19
Topic attachments
I Attachment History Action Size Date Who Comment
PNGpng RecoAllAssoc2GenMatched_PullX_vs_Pt_Sigma.png r1 manage 46.9 K 2022-10-20 - 09:59 GiorgioPizzati  
PNGpng RecoAllAssoc2GenMatched_PullZ_vs_Pt_Sigma.png r1 manage 48.1 K 2022-10-20 - 09:59 GiorgioPizzati  
PNGpng RecoAllAssoc2GenMatched_ResolX_vs_Pt_Sigma.png r1 manage 48.2 K 2022-10-20 - 09:59 GiorgioPizzati  
PNGpng RecoAllAssoc2GenMatched_ResolZ_vs_Pt_Sigma.png r1 manage 48.2 K 2022-10-20 - 09:59 GiorgioPizzati  
PNGpng TruePVLocationIndexCumulative.png r1 manage 36.9 K 2022-10-20 - 09:59 GiorgioPizzati  
PNGpng effic_vs_NumVertices.png r1 manage 44.7 K 2022-10-20 - 09:59 GiorgioPizzati  
PNGpng effic_vs_Pt2.png r1 manage 41.5 K 2022-10-20 - 09:59 GiorgioPizzati  
PNGpng fakerate_vs_PU.png r1 manage 45.2 K 2022-10-20 - 09:59 GiorgioPizzati  
PNGpng phase2_speedup.png r1 manage 33.5 K 2022-10-20 - 09:59 GiorgioPizzati  
PNGpng phase2_time.png r1 manage 33.0 K 2022-10-20 - 09:59 GiorgioPizzati  
Edit | Attach | Watch | Print version | History: r4 < r3 < r2 < r1 | Backlinks | Raw View | WYSIWYG | More topic actions
Topic revision: r4 - 2022-11-09 - MarcoMusich
 
    • Cern Search Icon Cern Search
    • TWiki Search Icon TWiki Search
    • Google Search Icon Google Search

    CMSPublic All webs login

This site is powered by the TWiki collaboration platform Powered by PerlCopyright & 2008-2023 by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
or Ideas, requests, problems regarding TWiki? use Discourse or Send feedback