Few months before the launch of the Run 2 reprocessing, the monolithic
MooSegmentFinderAlg which built the hough patterns and the segments was replaced by the
MuonSegmentFinderAlg and the
MuonLayerHoughAlg. That happened for two reasons: 1. Road finding & segment building a two very distinct steps. The breakup into two algorithms is more MT friendly. 2. The MuonSegmentFinderAlg is the only one that incorporates additionally NSW segment building. The switch was made at a time where the second NSW was a desirable option but still far away from its completion. A single algorithm was considered to ensure consistency between Run 2 & Run 3 geometry in the remaining MS. Enough of history, it turns out that slightly different code paths were taken in the same algorithm between the 2 setups anyway.
MuonSegmentFinder: At the very first, the PRDs are transformed into ROTs using the position & direction of the roads as initial seed. In the run 2 setup, the job was carried out by the
MuonPatternSegmentMaker, while in the other the
DCMathSegmentMakerwere used. Both were doing quite similar things but not exactly the same. The
MuonPatternSegmentMakeris replaced by the other two. The origins of the frozen-tier0-violating reference changes are pointed out in the Gitlab comments below.
NCB segments: Actually, the non-collision background segments were written into the Run 2 AODs while for the Run 3, none of them were implemented. Actually, I've no clue whether they're still needed or not. @jferrand, @christos, @mhodgkin, can you please ping the NCB group to confirm or deny that? In Run 2, the segments were nothing else than running CSC reconstruction but removing the IP constraint. Similar things are now realized with the NSW. Given that we know that the Micromega seeding will explode in terms of CPU time due to the huge cluster occupancy on each pad, these segments are only seeded from the sTGCs only. The Run 2 configuration is now updated to use the MuonSegmentMaker algorithim. Reference count changes in the
q442tests are verified to arise from the ordering of the segments within the container.
MuonLayerHoughAlg: In the monolithic version of the hough pattern finding, the
MuonHoughPatternFinderToolwas used in cosmic runs instead of the
MuonLayerHoughTool. No exchange of the tools was implemented for the
MuonLayerHoughAlg. Now fixed in the configuration. Btw. the garbage collector in the MS track building might lead to an explosion of the event recosntruction time. With explosion, I mean that I encountered events with hours and hours of processing time. May be !59990 (merged) should be merged before this one.
Checks were performed on 50k di muon events in the range 10-100 GeV using the Run 3 geometry
Full list of checks can be found here
Tagging: @giraudpf, @pscholer, @goblirsc, @atlasbot, @sangelid