Real-time physics: novel concepts for trigger, calibration & alignment, and data processing with LHCb Lucia Grillo on behalf of the LHCb Collaboration LHCP 2016 Lund, 13 - 18 June Picture: Horologium mirabile Lundense, the astronomical clock of Lund Cathedral 1
30
Embed
novel concepts for trigger, calibration & alignment, and data processing … · 2016-06-24 · Real-time physics: novel concepts for trigger, calibration & alignment, and data processing
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Real-time physics: novel concepts for trigger, calibration & alignment,
and data processing with LHCb
Lucia Grillo on behalf of the LHCb Collaboration
LHCP 2016Lund, 13 - 18 June
Picture: Horologium mirabile Lundense, the astronomical
Calorimeter System• Electromagnetic (ECAL) and hadronic
(HCAL) calorimeters
• Scintillator planes + absorber material planes
• Used the hardware (L0) trigger selection
• 5 stations equipped with multi-wire proportional chambers• Inner part of the first station equipped with GEM detectors• Used the hardware (L0) trigger selection
Performance of the Muon Identification system JINST 8 (2013) P10020
LHCb Detector Performance; Int. J. Mod. Phys. A30 (2015) 1530022
• Preliminary alignment and calibration of the detector
Trigger (online reconstruction)
Offline reconstruction
• Faster but less performing track and PID reconstruction (only part of the PID information available)
• Full and best performing detector alignment and calibration
• Full reconstruction including full PID information
Trigger != Offline
2012: Introduction of the Deferred Trigger
• Better usage of the farm exploiting the time between physics fillsPerformance of the LHCb high level trigger in 2012,
J. Phys. Conf. Ser. 513 (2014) 0120019
Trigger - Run I face
• Preliminary alignment and calibration of the detector
Trigger (online reconstruction)
Offline reconstruction
• Faster but less performing track and PID reconstruction (only part of the PID information available)
• Full and best performing detector alignment and calibration
• Full reconstruction including full PID information
Trigger != Offline
2012: Introduction of the Deferred Trigger
• Better usage of the farm exploiting the time between physics fillsPerformance of the LHCb high level trigger in 2012,
J. Phys. Conf. Ser. 513 (2014) 012001
Novel strategy: physics on the HLT output
• No need for offline data processing (the best possible selection of signal candidates is already there)
• Smaller event size allows to reduce pre-scales
more physics with the given resources
10
Trigger - Run II face
• All events are buffered to disk before running second software level of trigger
• Perform calibration of PID detectors and alignment of the full tracking system in real-time
• Last trigger level runs the offline-quality reconstruction
Trigger == Offline
➜ Same alignment and calibration constants in the trigger and offline
➜ The same full reconstruction in the trigger and offline: the best performance
11
Trigger - Run II face
• All events are buffered to disk before running second software level of trigger
• Perform calibration of PID detectors and alignment of the full tracking system in real-time
• Last trigger level runs the offline-quality reconstruction
Trigger == Offline
➜ Same alignment and calibration constants in the trigger and offline
➜ The same full reconstruction in the trigger and offline: the best performanceNovel strategy: physics on the HLT output
Challenges:
• Offline quality reconstruction in few ms
• Full detector alignment and calibration in real time
12
Alignment and calibration: how important!
Physics performance relies on spatial alignment and calibration of the detector
• Examples: VELO alignment is essential for PV discrimination, IP and proper time resolution,
First σIP (high pT ) = 14.0 µm
Latest σIP (high pT ) = 11.6 µm
13
Alignment and calibration: how important!
Physics performance relies on spatial alignment and calibration of the detector
LHCb Preliminary
LHCb Preliminary
• Examples: VELO alignment is essential for PV discrimination, IP and proper time resolution, better alignment improves mass resolution,
First alignment σΥ = 92 MeV/c2
Latest alignment σΥ = 49 MeV/c2
First σIP (high pT ) = 14.0 µm
Latest σIP (high pT ) = 11.6 µm
14
Alignment and calibration: how important!
Physics performance relies on spatial alignment and calibration of the detector
LHCb Preliminary
LHCb Preliminary
• Examples: VELO alignment is essential for PV discrimination, IP and proper time resolution, better alignment improves mass resolution, use of PID allows for more exclusive selections
Trigger+PID+tighter PID
Trigger+PID
CS: D+→π+π-π+
DCS: D+→K+-K-K+
First σIP (high pT ) = 14.0 µm
Latest σIP (high pT ) = 11.6 µm
First alignment σΥ = 92 MeV/c2
Latest alignment σΥ = 49 MeV/c2
15
Real-time alignment & calibration
• Alignments: VELO, Trackers, RICH mirrors, Muon
• Calibrations: RICH refractive index and HPDs, OT time, Calorimeters
16
Alignment & calibration framework• Strategy:
• Automatic evaluation at regular intervals (per run or per fill depending on the task)
• Dedicated sample to perform alignment or calibration collected with a specific trigger selection
• Compute new alignment or calibration constants (few minutes for the alignment tasks, run at the beginning of each fill when the VELO is closed)
• Update the constants if necessary
• New constants will be used in the trigger and coherently in offline reconstruction
2012: no correction 2015: applied new π0 calibration applied new LED corrections
• Absolute calibrations
• π calibration for ECAL: compute di-photon invariant mass.
• Cs source scan for HCAL, evaluated during the technical stop
0
22
Optimization of reconstruction
• Code optimization (vectorization, memory access)• New reconstruction chain in HLT1 and simplified geometry in Kalman filter
• Re-implementation and/or re-tuning of the algorithms (HLT1 and HLT2)• Run I offline reconstruction 2 times faster
Offline Run I reconstruction can run in the Trigger in Run II, with same or better performance
• New resources: farm nearly doubled wrt Run I, 10PB disk space to buffer the events between HLT1 and HLT2
• Big effort in speeding up the reconstruction:
Callgrind graphs(Area ∝ CPU time)
23
Run II performance: Tracking and PID
Same or improved performance as offline in Run I, but directly in the trigger
IP resolution
Track reconstruction efficiency
Pion misidentification efficiency vs Kaon identification efficiency
24
Run-II performance: Trigger
2015
2012
LHCb Preliminary
Efficiency for B+→D0π+ is ~75% Efficiency for B+→D0π+ is >90%
2011
Efficiency of the HLT2 inclusive beauty trigger as a function of B pT
• Improvement of the trigger efficiency thanks to e.g.• Best reconstruction already in the trigger• Detector fully aligned and calibrated• Use of PID information
2015
LHCb Detector Performance; Int. J. Mod. Phys. A30 (2015) 1530022
• Part of the physics programme needs billions of recorded candidates (e.g. charm measurements): but with no need for the rest of the event.
➜ “Turbo” stream
• Reduced event size (5kB vs 70kB)• No offline processing• ~2.5Hz of the output stream
(10kH for full)
Analyses can be done directly on the trigger output
26
Turbo and Turbo++
Even more analyses can be done on the trigger output
new for 2016
K
PV
D
!+
-
0
!+D*+
• Allows other reconstructed objects from the event to be saved, in addition to those selected by the trigger
New features:
• Saves only objects selected by the trigger
• Output limited to a standard set of variables
• Allows to create and save new variables (i.e. hits in a cone region around the track)
• Aim: according to the physics channel and desired measurement, choose how much (and which variables) of the event need to be saved
Out of the 420 HLT2 lines in 2016 physics programme, 150 choose Turbo, ~60 new lines wrt 2015
Turbo candidate
K
D
!+
-
0
!+PV *+D
Tracks from others PVs Other tracks
from trigger PV
+γ, π0Turbo++ candidate
27
Real-time physics: very first Turbo stream results
Few weeks after taking data
]2c [MeV/ -µ+µm2950 3000 3050 3100 3150 3200
2 cC
andi
date
s per
5 M
eV/
2
4
6
8
10
12
310!
-1 =3.05 pbintL = 13 TeV, sLHCb
c < 3 GeV/T
p2 < < 3.5y3 <
m(K−π+π+) [MeV/c2]1850 1900
Can
did
ate
s/(1
MeV/c
2)
0
50
100
×103
D+
Fit
Sig. + Sec.
Comb. bkg.
LHCb√s = 13TeV
m(K−π+) [MeV/c2]1800 1850 1900
Can
did
ate
s/(1
MeV/c
2)
0
50
100
150
×103
D0
Fit
Sig. + Sec.
Comb. bkg.
LHCb√s = 13TeV
28
Conclusions
• New data taking strategy implemented at LHCb experiment for Run II
• Detector calibration and alignment are provided in few minutes
• Best performance of the detector and full reconstruction, including PID, are exploited already in the trigger selection
• Thanks to the Turbo stream we are able to increase our physics program using the same resources, and to analyze data ~24h after having taken them
• We keep improving and speeding up the reconstruction, alignment, calibration and data processing
• Next challenge: upgrade!
• New development for 2016: Turbo++, to choose how much, and which variables of the events to be computed and saved according to the physics measurement