Performance of Run 3 track reconstruction with the mkFit algorithm

Results have been published in the detector performance note CMS-DP-2022-018

Track Reconstruction at CMS

This note reports on the CMS track reconstruction at the LHC Run 3.

In Run 2, the CMS track reconstruction algorithm used an iterative approach based on combinatorial Kalman Filter (CKF), consisting of twelve main iterations targeting different track topologies and seeded with different seed tracks. The name assigned to each iteration stems either from the order of execution, or from the seeding topology.

For Run 3, a new algorithm has been developed for track pattern recognition (or track building), named mkFit [1], that maximally exploits parallelization and vectorization in multi-core CPU architectures. This algorithm has been deployed in the CMS software for a subset of six tracking iterations:

• InitialPreSplitting, initial iteration before splitting merged pixel clusters in dense jet environments;
• Initial, initial iteration;
• HighPtTriplet, high- triplet iteration;
• DetachedQuad, detached quadruplet iteration;
• DetachedTriplet, detached triplet iteration;
• PixelLess, pixel-less iteration.

The mkFit algorithm allows to retain a similar physics performance with respect to the traditional CKF-based pattern recognition, while improving the computational performance of the CMS track reconstruction.

Tracking efficiency, fake rate and duplicate rate

The reconstructed tracks are selected with "high purity" quality flag [2].

A reconstructed track is considered associated to a simulated particle if more than 75% of its hits have been originated from this simulated particle. If this is not the case, the reconstructed track is considered as a random combination of hits and marked as a misidentified (fake) track.

Simulated tracks coming from the signal (hard scattering) vertex are used in the efficiency computation.
The tracking efficiency is defined as the fraction of simulated tracks associated to at least one reconstructed track.

All simulated tracks coming from any vertex (including pileup vertices) are used in the fake rate and duplicate rate computation.
The tracking fake rate is defined as the fraction of misidentified reconstructed tracks.
The tracking duplicate rate is defined as the fraction of reconstructed tracks associated multiple times to the same simulated track.

The performance has been measured in a simulated sample with superimposed pileup (PU) events. The number of PU events is sampled from minimum bias events with Poisson mean uniformly distributed from 55 to 75. The detector conditions are simulated with no module failure and account for the residual radiation damage due to Run 2 operations.

Figure	Description
pdf png	Tracking efficiency per iteration vs : The tracking efficiency per iteration is shown as a function of the simulated track , for simulated tracks with || <3.0 and || <2.5 cm. The mkFit algorithm is used for track pattern recognition in a subset of six tracking iterations: InitialPreSplitting, initial iteration before splitting merged pixel clusters in dense jet environments; Initial, initial iteration; HighPtTriplet, high- triplet iteration; DetachedQuad, detached quadruplet iteration; DetachedTriplet, detached triplet iteration; PixelLess, pixel-less iteration. The InitialPreSplitting iteration is only used to define tracking regions for the JetCore iteration, that targets track reconstruction within high- jets. All iterations except for InitialPreSplitting make use of split clusters. Contact: Mario Masciovecchio
pdf png	Tracking efficiency vs : The tracking efficiency is shown as a function of the simulated track for the CKF tracking (red) and the Run 3 tracking using mkFit (black), for simulated tracks with || <3.0 and || <2.5 cm. The tracking efficiency when mkFit is used for track building in a subset of six tracking iterations is consistent with the one obtained with the traditional CKF tracking algorithm. Contact: Mario Masciovecchio
pdf png	Tracking efficiency vs : The tracking efficiency is shown as a function of the simulated track pseudorapidity for the CKF tracking (red) and the Run 3 tracking using mkFit (black), for simulated tracks with >0.9 GeV and || <2.5 cm. The tracking efficiency when mkFit is used for track building in a subset of six tracking iterations is consistent with the one obtained with the traditional CKF tracking algorithm. Contact: Mario Masciovecchio
pdf png	Tracking efficiency vs PU: The tracking efficiency is shown as a function of the event pileup (PU) for the CKF tracking (red) and the Run 3 tracking using mkFit (black), for simulated tracks with >0.9 GeV, || <3.0 and || <2.5 cm. The tracking efficiency when mkFit is used for track building in a subset of six tracking iterations is consistent with the one obtained with the traditional CKF tracking algorithm. Contact: Mario Masciovecchio

pdf png	Tracking fake rate vs : The tracking fake rate is shown as a function of the reconstructed track for the CKF tracking (red) and the Run 3 tracking using mkFit (black). The tracking fake rate when mkFit is used for track building in a subset of six tracking iterations is on average lower than the one obtained with the traditional CKF tracking algorithm. Contact: Mario Masciovecchio
pdf png	Tracking fake rate vs : The tracking fake rate is shown as a function of the reconstructed track pseudorapidity for the CKF tracking (red) and the Run 3 tracking using mkFit (black), for tracks with >0.9 GeV. The tracking fake rate when mkFit is used for track building in a subset of six tracking iterations is on average lower than the one obtained with the traditional CKF tracking algorithm. Contact: Mario Masciovecchio
pdf png	Tracking fake rate vs PU: The tracking efficiency is shown as a function of the event pileup (PU) for the CKF tracking (red) and the Run 3 tracking using mkFit (black), for tracks with >0.9 GeV. The tracking fake rate when mkFit is used for track building in a subset of six tracking iterations is on average lower than the one obtained with the traditional CKF tracking algorithm by about 0.5%. Contact: Mario Masciovecchio

pdf png	Tracking duplicate rate vs : The tracking duplicate rate is shown as a function of the reconstructed track for the CKF tracking (red) and the Run 3 tracking using mkFit (black). The tracking duplicate rate when mkFit is used for track building in a subset of six tracking iterations is higher than the one obtained with the traditional CKF tracking algorithm at 0.5 <<2 GeV, while it's lower or equal elsewhere. Contact: Mario Masciovecchio
pdf png	Tracking duplicate rate vs : The tracking duplicate rate is shown as a function of the reconstructed track pseudorapidity for the CKF tracking (red) and the Run 3 tracking using mkFit (black), for tracks with >0.9 GeV. The tracking duplicate rate when mkFit is used for track building in a subset of six tracking iterations is higher than the one obtained with the traditional CKF tracking algorithm especially at 1.45<||<2.5, while it's lower at ||>2.5. Contact: Mario Masciovecchio
pdf png	Tracking duplicate rate vs PU: The tracking duplicate rate is shown as a function of the event pileup (PU) for the CKF tracking (red) and the Run 3 tracking using mkFit (black), for simulated tracks with >0.9 GeV. The tracking duplicate rate when mkFit is used for track building in a subset of six tracking iterations is on average higher than the one obtained with the traditional CKF tracking algorithm by about 0.5%. Contact: Mario Masciovecchio

Time performance

The tracking time performance has been measured in the same simulated sample with superimposed pileup (PU) events as for the physics performance. The number of PU events is sampled from minimum bias events with Poisson mean uniformly distributed from 55 to 75.
Single-threaded measurements are performed with local access to the input.

Figure	Description
pdf png	Time performance: all iterations The tracking time is shown as a function of the tracking steps for the CKF tracking (red) and the Run 3 tracking using mkFit (black), for all tracking iterations. The vertical (time) scale is normalized to have the total time without mkFit equal to unity. Overall, using mkFit in a subset of six tracking iterations allows to reduce the track building time by a factor of about 1.7, corresponding to a reduction of the total tracking time by about 25%. In Run 3, tracking has been measured to make about half of the total offline reconstruction time. Thus, this translates to a reduction of the total offline CMS reconstruction time or conversely to an increase of the event throughput by 10-15%. Contact: Mario Masciovecchio
pdf png	Time performance: mkFit iterations The tracking time is shown as a function of the tracking steps for the CKF tracking (red) and the Run 3 tracking using mkFit (black), for the subset of six tracking iterations using mkFit for track building. The vertical (time) scale is normalized to have the total time without mkFit for all tracking iterations equal to unity. Using mkFit allows to reduce the track building time by a factor of about 3.5 considering the sum of six iterations where mkFit is employed. In individual iterations where mkFit is employed, this factor varies from about 2.7 to about 6.7. Contact: Mario Masciovecchio

References

1. Speeding up particle track reconstruction using a parallel Kalman filter algorithm, Steven Lantz (Cornell U.), Kevin McDermott (Cornell U.), Michael Reid (Cornell U.), Daniel Riley (Cornell U.), Peter Wittich (Cornell U.) et al., e-Print: 2006.00071 [physics.ins-det], DOI: 10.1088/1748-0221/15/09/P09030, Published in: JINST 15 (2020) 09, P09030.
2. Description and performance of track and primary-vertex reconstruction with the CMS tracker, CMS Collaboration, e-Print: 1405.6569 [physics.ins-det], DOI: 10.1088/1748-0221/9/10/P10009, Published in: JINST 9 (2014) 10, P10009.