News from GPU parameters tuning
Single kernel optimization
MergerTrackFit




MergerSectorRefit


MergerCollect



CompressionStep1unattached



4 dimensions optimization (MergerFollowLoopers + CompressionStep0attached)


21 dimensions optimization - SectorTracker step





Clusterizer step


Automated tuning
- Developed script for automated tuning
- Tunes most of the steps
- Tried on a 750kHz pp simulated dataset
- Results are for the sync time of the standalone benchmark
- Sync mean time default: 1440.69 ms ± 3.94 ms
- Sync mean time optimised: 1318.73 ms ± 4.97 ms
- Performance gain 8.47%
Next two weeks
- Absence for CERN School of Computing
- Plan to run some more automated tuning (on more datasets)
- Create collection of parameter dumps