slowdown in PFCandidatePrimaryVertexSorter in CMSSW_12_1_0_pre2 -> CMSSW_12_1_0_pre3

[jpata]

Slava noticed that it looks like the PFCandidatePrimaryVertexSorter is called 4x in CMSSW_12_1_0_pre3 [1] instead of 2x in CMSSW_12_1_0_pre2 [2], and it's a fairly significant increase in the step4 runtime.

There are two new callers:

• primaryVertexAssociationJME
• primaryVertexAssociationJMEnoHF

[1] http://cms-reco-profiling.web.cern.ch/cms-reco-profiling/circles/piechart.php?local=false&dataset=CMSSW_12_1_0_pre3%2Fslc7_amd64_gcc900%2F11834.21%2Fstep4_circles&resource=time_thread&colours=default&groups=packages&threshold=0

[2] http://cms-reco-profiling.web.cern.ch/cms-reco-profiling/circles/piechart.php?local=false&dataset=CMSSW_12_1_0_pre2%2Fslc7_amd64_gcc900%2F11834.21%2Fstep4_circles&resource=time_thread&colours=default&groups=packages&threshold=0

@laurenhay @ahinzmann could this be related to #33885?

[jpata]

assign reconstruction

[cmsbuild]

New categories assigned: reconstruction

@slava77,@jpata you have been requested to review this Pull request/Issue and eventually sign? Thanks

[cmsbuild]

A new Issue was created by @jpata Joosep Pata.

@Dr15Jones, @perrotta, @dpiparo, @makortel, @smuzaffar, @qliphy can you please review it and eventually sign/assign? Thanks.

cms-bot commands are listed here

[slava77]

Looking at https://cms-reco-profiling.web.cern.ch/cms-reco-profiling/cgi-bin/igprof-navigator/releases/CMSSW_12_1_0_pre4/34834.21/step4/cpu_endjob/21
or https://cms-reco-profiling.web.cern.ch/cms-reco-profiling/cgi-bin/igprof-navigator/releases/CMSSW_12_1_0_pre4/11834.21/step4/cpu_endjob/22

it's a bit puzzling to see that a call to chargedHadronVertex(vertices, iVertex, ... does not appear in the cost stack, as if it was completely inlined in the chargedHadronVertex(vertices, trackRef...
(I doubt that its cost is so small that it would never show up in the sampling).

Perhaps for debugging it's worth to compile the two functions in two different .cc files?

[ahinzmann]

yes, this is because the new modules primaryVertexAssociationJME is taking the track-vertex-association-decision for CHS and PUPPI. This could be speed up, by not reusing the same module, but instead making a more light-weight association module for CHS and PUPPI. We plan to implement this, once the CHS/PUPPI vertex-association criteria for Run3 are fixed.

[jpata]

We plan to implement this, once the CHS/PUPPI vertex-association criteria for Run3 are fixed.

could you clarify a bit what's the timeline for this?

[ahinzmann]

so first studies are done, but need to iterate order of 2 weeks, before taking a decision

[jpata]

I believe this was addressed in #36639.

[slava77]

Slava noticed that it looks like the PFCandidatePrimaryVertexSorter is called 4x in CMSSW_12_1_0_pre3 [1] instead of 2x in CMSSW_12_1_0_pre2 [2], and it's a fairly significant increase in the step4 runtime.

looking at http://cms-reco-profiling.web.cern.ch/cms-reco-profiling/circles/piechart.php?local=false&dataset=CMSSW_12_3_0_pre4%2Fslc7_amd64_gcc10%2F11834.21%2Fstep4_circles&resource=time_thread&colours=default&groups=packages&threshold=0

I still see 4 cases taking about 13% of CPU time; compared to 12% in CMSSW_12_1_0_pre3 when the issue was opened.
Am I missing something?