Open clacaputo opened 1 year ago
assign reconstruction
New categories assigned: reconstruction
@mandrenguyen,@clacaputo you have been requested to review this Pull request/Issue and eventually sign? Thanks
A new Issue was created by @clacaputo Claudio Caputo.
@Dr15Jones, @perrotta, @dpiparo, @rappoccio, @makortel, @smuzaffar can you please review it and eventually sign/assign? Thanks.
cms-bot commands are listed here
From a first investigation on the reconstruction-related PRs integrated into pre3
, there is no clear responsible
From a first investigation on the reconstruction-related PRs integrated into
pre3
, there is no clear responsible
the summary for the data wf seems to clearly point to pfParticleNetFromMiniAODAK4CHS{Central,Forward}*
It may be clarified in the issue description; now it sounds like the origins are similar.
apparently your workflow redoes gensim from scratch. I see some difference in the number of sim tracks but at least the mean is within 0.2% (red is pre2; plot is from step1.root on eos)
it looks like minbias has BeamSpotObjects_Realistic25ns_13p6TeVCollisions_EOY2022_v1_mc while the pre3 GT has BeamSpotObjects_Realistic25ns_13p6TeVCollisions_EOY2022_v2_mc. The changes are relatively minor (v1 has a wider y), ~but eventually something to get in sync.~ (this is intentional/unavoidable and is OK, apparently) This change happened from pre2->pre3
https://github.com/cms-sw/cmssw/releases/CMSSW_13_1_0_pre3 has changes in geant
Can you run another variant of pre3 using step1 inputs from pre2 as well as using 131X_mcRun3_2022_realistic_v1? to decouple beam spot and geant effects.
This change happened from pre2->pre3
the change in beam spot in the GT since pre2 is actually more significant than the v1 and v2 I mentioned above for minBias (the minbias in pre3 is consistent with the pre3 GT).
Can you run another variant of pre3 using step1 inputs from pre2 as well as using 131X_mcRun3_2022_realistic_v1? to decouple beam spot and geant effects.
and older minBias as well
Can you run another variant of pre3 using step1 inputs from pre2 as well as using 131X_mcRun3_2022_realistic_v1? to decouple beam spot and geant effects.
Hi @slava77 , thanks for the suggestion. Talking with @mandrenguyen we arrived at the same conclusion. I'll run it
the change in beam spot in the GT since pre2 is actually more significant than the v1 and v2 I mentioned above for minBias (the minbias in pre3 is consistent with the pre3 GT).
for the record, here's the change.
mainly moving the z coordinate of the ellipsoid center towards 0 and related uncertainty. Descriptions at https://github.com/cms-sw/cmssw/pull/41123#issue-1634394648
Can you run another variant of pre3 using step1 inputs from pre2 as well as using 131X_mcRun3_2022_realistic_v1? to decouple beam spot and geant effects.
I did the check running pre3
using step1 inputs from pre2 as well as using
131X_mcRun3_2022_realistic_v1`
Timing:
pre2
: 12.23 s/evpre3
: 13.22 s/evpre3_with_pre2BS
: 12.32 s/evModules' time differences pre2
VS pre3_with_pre2BS
The same excluding the first 1 events
delta/mean delta/orJob original new module name
---------- ------------ -------- ---- ------------
-0.051121 -0.06% 152.05 ms/ev -> 144.47 ms/ev trackExtrapolator
+0.050967 +0.23% 517.27 ms/ev -> 544.32 ms/ev tobTecStepTrackCandidates
---------- ------------ -------- ---- ------------
It is clear that the reconstruction time increase in pre3
is coming from the new BeamSpot conditions, introduced by #41123
Hi @clacaputo
It is clear that the reconstruction time increase in pre3 is coming from the new BeamSpot conditions, introduced by https://github.com/cms-sw/cmssw/pull/41123
Marco pointed me to this issue: indeed the new BeamSpot was introduced because it reflects better the Z position of the Beamspot in the 2022 data (see this presentation for a few more details). So the new measure is correct. Interesting that 1 cm in the BeamSpot Z position has such a large effect on reco-timing!
Just to be sure we (me & Marco) took a look at the relvals introduced in the same PR to mach the new BeamSpot ('CMSSW_13_1_0_pre1-130X_mcRun3_2022_realistic_withNewBSFromEOY2022Data_v2_RV186-v'
) and we concluded that they were correctly generated:
--beamspot Realistic25ns13p6TeVEOY2022Collision
)with a preliminary BeamSpot tag in the GT (BeamSpotObjects_Realistic25ns_13p6TeVCollisions_EOY2022_v1_mc
) which is compatible with the final measurement (*see attached screenshot)
Given that the increase shows most prominently in the TobTec step, which I guess sort of makes sense for a shift in the z coordinate of the beamspot, we probably ought to check whether the fraction of tracks being reco'd and selected with highPurity in that iteration is changing.
we probably ought to check whether the fraction of tracks being reco'd and selected with highPurity in that iteration is changing.
that's precisely what I wanted to check on the RelVal samples, while on paper everything seems to be correct, I am a bit puzzled by the results in the release validation. I checked:
/RelValTTbar_14TeV/CMSSW_13_1_0_pre3-131X_mcRun3_2022_realistic_v2-v1/DQMIO
(pre3, black)/RelValTTbar_14TeV/CMSSW_13_1_0_pre2-131X_mcRun3_2022_realistic_v1-v2/DQMIO
(pre2, blue)
and while the beamspot is shifted as expected (https://tinyurl.com/2aa3b6fr) there is a degradation of performance:
The same thing is observed in the samples with PU:
RelValTTbar_14TeV/CMSSW_13_1_0_pre3-PU_131X_mcRun3_2022_realistic_v2-v1/DQMIO
(pre3, black)/RelValTTbar_14TeV/CMSSW_13_1_0_pre2-PU_131X_mcRun3_2022_realistic_v1-v1/DQMIO
(pre2, blue)performance:
I would tend to conclude that the samples in the RelVals were generated with recycled GEN-SIM (which in the case at hand doesn't make sense :( ).
I did the check running
pre3
using step1 inputs frompre2 as well as using
131X_mcRun3_2022_realistic_v1`
did you use the pre2 minbias GEN-SIM as well?
did you use the pre2 minbias GEN-SIM as well?
~Shouldn't be done automatically using pre2 step2
as input of pre3 step3
?~
I'm running pre3 step3
again
~Shouldn't be done automatically using
pre2 step2
as input ofpre3 step3
?~
it's probably good enough in this case (only BS changed in GT and hopefully digi formats didn't change)
~Shouldn't be done automatically using
pre2 step2
as input ofpre3 step3
?~it's probably good enough in this case (only BS changed in GT and hopefully digi formats didn't change)
pre3 step3
tested with:
Time measured: 12.4 s/ev
Profiling
11834.21
inCMSSW_13_1_0_pre3
, we measured an overall increase of ~1 secs in the reconstruction timing with respect toCMSSW_13_1_0_pre2
. The measure has been performed onvocms011
averaging on 5 different measurements:Investigating the modules' time differences doesn't provide any insight, but it seems a general time increase
Investigating
igprof
output shows a 1% increase incms::CkfTrackCandidateMakerBase::produceBase
as you can see here.NOTA: For
136.889
(T0-like condition on Data), the time increase is not so high8.09688 s/ev ==> 8.34514 s/ev
link and it is clearly related topfParticleNetFromMiniAODAK4CHS{Central,Forward}*