Tracking recent developments and activities

Speaker : Jason Webb ( BNL )


Talk time : 12:00, Duration : 00:20

Tracking Code for 2015 data production

  • Need to make a decision on tracker for 2015 data calibration and production
    • Sti is the baseline choice
    • Stv has advanced to the point (QA) where we can consider it, but speed still an issue
  • Investigated performance (speed) of Stv relative to Sti with HFT in tracking
    • Stv takes about 2.6x as long to perform track reconstruction as Sti
  • Significant speed gains for both Stv and Sti could be made if we can reject bad seeds early in the fit, since only 1 in 10 seeds result in a good fit track.
    • Stv Knn (k-Nearest Neighbor) seed finder investigated
    • Shows 10% improvement in speed, relative to standard Stv
      • Further gains possible as Knn can be used to propagate the track, reducing number of geometry calls.  (Requires further development beyond current scope).
    • Stv still takes 2.4x as long as Sti to perform track reconstruction
  • Decision is to go with Sti for the 2015 data calibration and production
  • Knn may provide increased speed for Sti
    • Preliminary tests (low statisitics) show 6% gain in speed. 
    • If this holds up with more rigouros test, will evaulate QA and possibly tune Sti/Knn.
    • Decision for seed finder will be made soon
  • Evaluating DEV vs SL15e/g production library
    • Changes in DEV since SL15e/g:
      • Additional refit added to improve propagation of track from inner to outer, to better pick up outermost hits
      • Tree Search algo rewritten in order to eliminate small (1/10k) fraction of tracks with insuffficient number of HFT hits
    • QA performed with TPC only tracking shows consistent results
    • QA performed with HFT in tracking to be done.  Currently debugging simulation chain.

Integration of SST into tracking (ongoing)

  • Experience with integrating IST & PXL (and memory of the SVT times) suggests that we need to have multiple layers of the HFT participating in the track, otherwise probablitiy of picking up a bad (but high precision) hit is high
  • Will evaluate 3-layer combinations:
    • PXL1 * PXL2 * IST -- baseline
    • PXL1 * PXL2 * IST * SST -- most restrictive
    • (PXL1 * PXL2) * (IST + SST) -- require inner two layers
    • (PXL1 + PXL2) * (IST * SST) -- require outer two layers
    • PXL1 * (PXL2 + IST) * SST -- require inner-most and outer-most layers
  • Currently hung up debugging simulation chain