Geant4 Acceptance Suite for Key Observables CHEP06, T.I.F.R. Mumbai, 13-17 February 2006 J. Apostolakis, I. MacLaren, J. Apostolakis, I. MacLaren, P. Mendez.
Post on 31-Dec-2015
216 Views
Preview:
Transcript
Geant4 Acceptance SuiteGeant4 Acceptance Suitefor Key Observables for Key Observables
CHEP06, T.I.F.R. Mumbai, 13-17 February 2006
J. Apostolakis, I. MacLaren, J. Apostolakis, I. MacLaren,
P. Mendez Lorenzo, A. Ribon, J.P. P. Mendez Lorenzo, A. Ribon, J.P. WellischWellisch
CERN PH/SFTCERN PH/SFT
Alberto Ribon, CERN/PH/SFT
2
•
Motivation and strategy Description of the suite Some results for the December 2005
Geant4 release (G4 8.0)
Another application Summary and outlook
OutlineOutline
Alberto Ribon, CERN/PH/SFT
3
Our goal is to detect any difference between two Geant4 versions, by comparing a large number of physics observables.
Only those distributions which are statistically different (i.e. unlikely to be originated from the same parent distribution, whatever it is) should be visually examined by someone, to understand the origin of such difference, either a bug or an improvement.
As a start, we consider only calorimeter observables sensitive to hadronic physics.
Motivation and strategyMotivation and strategy
Alberto Ribon, CERN/PH/SFT
4
It is very important that the regression testing is automatic , because only in this way it is possible, in practice, to consider a very large number of distributions, and use it regularly, at least before each new release.
To do that we need a statistical package that offers several statistical tests (not only Chi2 and Kolmogorov-Smirnov!), to increase the chances to spot any kind of discrepancy between two distributions.
Some remarksSome remarks
Alberto Ribon, CERN/PH/SFT
5
It reproduces, in a simplified way, all the LHC calorimeters: Fe-Sci , Cu-Sci , Cu-LAr , W-LAr
Pb-Sci , Pb-LAr , PbWO4 .
Beam particle type: ± , K± , KL0 , p , n , e- .
Beam energy: 1, 2, 3, …, 10, 20, 30, 40, 50, 60, 80, 100, 120, 150, 180, 200, 250, 300 (1000) GeV.
The calorimeter is a box. The user can choose: - the total thickness of the absorber (in [mm] or ) - the thickness of the active layer - the number of layers - the number and the size (in [mm] or ) of the rings for the lateral shower profile.
Simplified Calorimeter setup
Alberto Ribon, CERN/PH/SFT
7
Observables: - total energy deposit in all active layers - total energy deposit in the whole calorimeter - energy deposit in each active layer (longitudinal shower profile) - energy deposit in each ring (i.e. radial bin) (lateral shower profile). The program produces in output a HBOOK n-
tuple, which stores all the above distributions. Then another program reads two of these n-
tuples, makes the statistical tests of the distributions, and produces a .PS file whenever the p-value of a test is below a certain threshold (currently 1% ).
Alberto Ribon, CERN/PH/SFT
8
7 calorimeters x 8 particles x 24 beam energies
x 5000 events x 5 Geant4 physics configurations.
About ≥ 0.07 sec/GeV ( @1 GHz CPU) to simulate a hadronic shower.
Use some biasing techniques (keep 1 e-/e+//n every 2 or 4 with weight 2 or 4), and kill neutrons below 1 MeV, to speed up the execution.
Overall, it takes few years of CPU time, but concentrated in 1-2 weeks: need of the GRID !
CPU issue
Alberto Ribon, CERN/PH/SFT
9
G4 8.0 candidate versus G4 7.1.p01 .
10 lambdas, 20 layers, 4 mm active layers, 10 rings of increasing size, starting from 0.1 lambdas; biasing for beam energies 10
GeV .
SLC 3.0.x g++ 3.2.3 . Run on the GRID as Geant4 Virtual
Organization. ≈ 120 CPUs in 7 sites in 6 countries.
≈1300 jobs per each Geant4 physics configuration.
Each job that ends normally produced a tar-ball
of about 3 MB size containing the results.
Geant4 release 8.0 (December 2005)
Alberto Ribon, CERN/PH/SFT
10
5 GRID installations; ≈ 12,000 jobs submitted to the GRID; ≈ 50,000 PostScript files produced (about 10%
of observables that have been checked)
≈ 99% GRID success rate (this is a significant improvements w.r.t.
previous GRID deployments, thanks to a careful selections of the sites where to run).
5 Geant4 physics configurations; 3 bugs found; 1 critical issue identified (elastic scattering).
Results
Alberto Ribon, CERN/PH/SFT
11
Log scale
G4 8.0 cand
G4 7.1.p01
Visible energy in the first ring
Visible energy in the first ring
P-value = 0.9%
Alberto Ribon, CERN/PH/SFT
12
Log scale
G4 8.0 cand
G4 7.1.p01
Total visible energy
Total visible energy
P-value = 10-18
Alberto Ribon, CERN/PH/SFT
13
As post-processing, using only the log files produced during the simulation, some summary .PS plots can be produced:
- longitudinal shower profiles - transverse shower profiles - sampling fractions vs. beam energy - energy resolutions vs. beam energy - e/ ratios vs. beam energy Some unphysical “jumps” are observed at 10
GeV, because biasing is not applied below such a value.
Summary plots
Alberto Ribon, CERN/PH/SFT
16
It is interesting and useful to study the effects of
physics changes (e.g. in multiple scattering, hadronic elastic scattering, total cross-sections, and in any other physics model), with a
simplified calorimeter setup, with pure MC information, without real life complications (e.g. beam composition/profile, noise, cross-talk,
clustering, reconstruction).
Another use of this tool
Alberto Ribon, CERN/PH/SFT
17
Studies in simplified calorimeter setups (I)
“CMS HCAL”- 150 GeV
Cu-Sci10 25 layers
(6 cm Cu)4 mm Sci
5000 eventsG4 7.1p01
QGSPLHEP
Log scale
Alberto Ribon, CERN/PH/SFT
18
Studies in simplified calorimeter setups (II)
“CMS HCAL”- 150 GeV
Cu-Sci10 25 layers
(6 cm Cu)4 mm Sci
5000 eventsG4 7.1p01
QGSPLHEP
Log scale
Alberto Ribon, CERN/PH/SFT
19
Summary
We have a tool which is useful for three purposes:
robustness testing;
regression testing;
detection of physics effects on calorimeter observables.
We are using it since one year, and we are planning to deploy it even more extensively in the future.
Alberto Ribon, CERN/PH/SFT
20
OutlookWe aim to improve and extend this tool in the
following directions: Include more statistical tests , and study their statistical power.
Find complementary (in particular, more global and faster) ways to spot differences.
Monitor CPU time, number of steps and tracks, and other observables.
Extend to electromagnetic shower shapes, and integrate all the efforts for testing Geant4.
top related