1. The LHCb Trigger Niko Neufeld CERN, PH
It is sometimes said that LHCb is a trigger and a RICH. Even if that’s an oversimplification it is a good pretence for this complementary talk at a RICH workshop

2. Outline The 3 levels of the LHCb Trigger
Level-0 hardware trigger
Fully synchronous and pipe-lined (deadtime < 0.5%)
Pile-up System
Calorimeter and Muon
Flexible L0 Decision unit
Level-1 software trigger
Partial read-out: Vertex Detector (VeLo), Trigger Tracker (TT) and L0 summary
High Level software trigger (HLT)
Full read-out: all detector data
Using the RICH in the High Level Trigger
Common
hardware
Visible cross section 100 p
Niko Neufeld
LHCb Trigger, RICH-2004

3. The LHCb trigger at a single glance
Niko Neufeld
LHCb Trigger, RICH-2004
~ O(1) kHz

4. Trigger Rates Overview
Level-0
30 MHz pp x-ings
10 MHz
Multiplicity/Pile-Up: 7 MHz
ET(m1,m2,h,e,g,p0): 1 MHz
Level-1
VELO: impact parameter
VELO+TT: momentum
VELO+L0-m: Mmm
HLT
L1(VELO+TT)+T: 10 kHz
VELO+TT+T: dp/p<1% exclusive channels, full reconstruction for 200 Hz
HLT
L1-confirmation
HLT
Full reconstruction
Level-0
Level-1
Niko Neufeld
LHCb Trigger, RICH-2004

5. Level-1 Decision Algorithm
PT1, PT2 µµ γ e
Parallel (overlapping) trigger lines
PT1, PT2?
PT1, PT2?
Bandwidth division:
L1PT
L1J/ψ
L1µµ
L1µ
L1γ
L1e
OR(L1) ⇒ L1 yes/no
Bandwidth (kHz)
Photon
Electron
Dimuon, J/Psi
Dimuon, general
Single-muon
Adjusted for overlap
Generic
30.0 (75.2%)
8.8 (22.1%)
1.7 ( 4.1%)
1.8 ( 4.6%)
3.9 ( 9.9%)
4.0 (10.0%)
3.0 ( 7.4%)
1.2 ( 3.0%)
1.1 ( 2.7%)
2.3 ( 5.9%)
2.3 ( 5.8%)
Overlaps are absorbed in this direction
Niko Neufeld
LHCb Trigger, RICH-2004

6. Efficiency for generic L1 trigger
Niko Neufeld
LHCb Trigger, RICH-2004

7. Key features of DAQ hardware
All detectors use standardized read-out boards (two variants  next slide)
Use commercial (mostly even commodity) components wherever possible: PCs, (copper) Gigabit Ethernet, Ethernet routers
Use the same infrastructure (network and computer farm) for L1 and HLT
Accommodate a soft real-time requirement: Level 1 latency no larger than 58 ms
Large system: 3000 Gigabit Ethernet links, 1800 PCs, several 100 Ethernet switches
Niko Neufeld
LHCb Trigger, RICH-2004

8. The RICH readout board Standardised read-out boards: 9U x 400 mmm
Gets data from detectors, from up to 48 optical links de-serialises, zero-suppresses, etc…
All boards are controlled by commercial Microcontroller (Creditcard-PC)
Data sent out by standard Gigabit Ethernet Mezzanine card via up to 4 Gigabit Ethernet links (over copper)
Niko Neufeld
LHCb Trigger, RICH-2004

9. Software trigger Hardware (DAQ)
Front-end Electronics
Level-1 Traffic
HLT Traffic FE FE FE FE FE FE FE FE FE FE FE FE
TRM
1000 kHz
5.5 GB/s
40 kHz
1.6 GB/s
Multiplexing Layer
Switch
Switch
Switch
Switch
Switch
TIER0
Readout Network
L1-Decision
Sorter
TFC System
Storage System
~ 250 MB/s
total
7.1 GB/s
SFC
Switch
CPU
SFC
Switch
CPU
SFC
Switch
CPU 94
SFCs
SFC
Switch
CPU
SFC
Switch
CPU
Scalable in depth: more CPUs
Scalable in width: more detectors in Level-1
CPU Farm
~1800 CPUs
Niko Neufeld
LHCb Trigger, RICH-2004

10. Using the RICH in the High Level Trigger

11. RICH in HLT: What does it get us & what does it cost us?
HLT involves exclusive selection of (many) B decay modes by checking the invariant-mass of combinations of 2–8 tracks eg Bs  Ds+Ds-  K+K-p+ K+K-p-
Long tracks (tracks having info from VeLo and T-stations) are used ~ 30 per event If only selection is on charge, then for Bs  Ds+Ds- ~ 156 combinations  107
If K/p identification were available, the number of combinations to be checked would be substantially reduced ~ 24 ×102  103 combinations for Bs  Ds+Ds- example  HLT could take less time per channel, if K/p ID is fast
10 ms/event to run HLT on a CPU in ms/event for exclusive selections
Niko Neufeld
LHCb Trigger, RICH-2004

12. First step: parameterise ring distortions so that problem can be solved on HPD planes avoiding the need to determine the Cherenkov angle for each pixel-track combination:
RICH-2
Hit pixels shown on HPD detector planes, for a typical event Crosses mark impact point of tracks (as if they were reflected)
Niko Neufeld
LHCb Trigger, RICH-2004

13. Same event: now ray trace “fake” photons to the detector plane, emitted from each track at fixed qC = 30 mrad, uniformly distributed around azimuthal angle f:
Niko Neufeld
LHCb Trigger, RICH-2004

14. Plot radius r vs f to calibrate the distortion of the rings
If plotted relative to the average radius r, all rings show  the same distortion: Dr = A cos 2f (A  2.5 mm)
Niko Neufeld
LHCb Trigger, RICH-2004

15. Reconstructed Cherenkov angle for all long tracks passing through RICH-2, vs momentum (on log scale)
Niko Neufeld
LHCb Trigger, RICH-2004

16. Same plot selecting out the true kaons only Note that below threshold, peak search finds ~ random qC
Niko Neufeld
LHCb Trigger, RICH-2004

17. Then apply correction: qC = 30 r / (r - A cos 2f) mrad
Determine the average ring radius for each track by ray tracing a few photons (currently 6) — fast
Then apply correction: qC = 30 r / (r - A cos 2f) mrad
Ray traced photons
Pixel hits from full simulation
s = 0.1 mrad
s = 0.7 mrad
 ~ as good as offline resolution on qC
Niko Neufeld
LHCb Trigger, RICH-2004

18. Single track All tracks
Plot reconstructed qC for all photons relative to a track
“Local” algorithm can be made by searching for peak, treating hits from other tracks as background
Scan over qC to find value that maximizes significance
Single track
All tracks
Niko Neufeld
LHCb Trigger, RICH-2004

19. Selecting the kaon band :
Cutting the pion band :
Selecting the kaon band :
RICH 2
RICH 2
Cut on (rS/B max-rtrue pion)
Cut on (rS/B max-rtrue kaon)
Niko Neufeld
LHCb Trigger, RICH-2004

20. Local HLT algorithms: performance
Kaon-ID efficiency (purple) / pion misid (blue) using gas info
Cut pion band
Select kaon band
Good performance & fast
Niko Neufeld
LHCb Trigger, RICH-2004

21. Global HLT algorithm: principle
Implement global LogLikelihood with some simplifications with
respect to full glory of “offline” reconstruction:
Cherenkov angle calculated on HPD plane (discussed)
Do not use aerogel (for the moment)
Do not calculate contribution of every pixel to every track;
rather assign pixels to ring image closest to track
Only consider pion vs kaon hypotheses in LL
Main advantages of this w.r.t. local approach:
Simultaneous treatment of signal and background
Method much better suited to looking for below threshold
kaons, which is a v. challenging problem for local method
Niko Neufeld
LHCb Trigger, RICH-2004

22. Global HLT algorithm: performance
Online
Offline
e(KK) = 88%
e(pK) = 15%
e(KK) = 92%
e(pK) = 11%
Good (but not exact) correlation with offline. Speed 11 ms (1 GHz PIII) per event when only “long” tracks are used. Promising!
Niko Neufeld
LHCb Trigger, RICH-2004

23. Summary LHCb uses a 3 level trigger system
Two levels of software trigger provide maximum flexibility at high rate
RICH information is available in the trigger and potentially very useful
Fast algorithms are being developed and look promising
We are currently implementing, deploying and eagerly awaiting 2007
Niko Neufeld
LHCb Trigger, RICH-2004

24. Acknowledgements
The work of many people has been presented in this talk
I would like thank in particular the LHCb RICH, Online and Electronics groups
Niko Neufeld
LHCb Trigger, RICH-2004

25. Backup Slides

26. Data Flow 1MHz Level 0 trigger Readout Network Switch Switch Switch
HLT Traffic
Level-1 Traffic
Front-end Electronics FE FE FE FE FE FE FE FE FE FE FE FE
TRM
323 Links
4 kHz
1.6 GB/s
126 Links
44 kHz
5.5 GB/s
Multiplexing
Layer
62 Switches
Switch
Switch
Switch
Switch
Switch
29 Switches
64 Links
88 kHz
32 Links
Readout Network
L1-Decision
Sorter
TFC System
Storage System
94 Links
7.1 GB/s
SFC
Switch
CPU
SFC
Switch
CPU
SFC
Switch
CPU 94
SFCs
SFC
Switch
CPU
SFC
Switch
CPU
CPU Farm
Gb Ethernet
Level-1 Traffic
Mixed Traffic
HLT Traffic
~1800 CPUs
Niko Neufeld
LHCb Trigger, RICH-2004

27. L1 efficiencies overview
Offline selected events:
(1.55 +/- 0.18)% - Why so low?
Reconstructible events:
Hadrons
trigger
e and !
4-prong give best generic efficiency!
Niko Neufeld
LHCb Trigger, RICH-2004

28. LHCb Software triggers
Level-1
HLT
Input rate
1 MHz
Stage 1) 40 kHz
Stage 2) 10 kHz
Data used
VeLo, TT, L0 summary
Stage 1) like L1
Stage 2) all
Output rate
40 kHz
O(2 kHz) out of which 200 Hz fully reconstructed
Mean time on 2007 CPU
1 ms
10 ms (total)
10 ms (Stage 2)
Maximum allowed time
58 ms (data transport included)
not applicable (limited only by available CPU power)
Niko Neufeld
LHCb Trigger, RICH-2004

29. Level-0: B-signatures m Global event variables: Trigger on B’s:
reject “complicated” events
Trigger on B’s:
thresholds on highest ET:
h-cluster
e-cluster
g-cluster
p0-cluster
HCAL clusters m
Or highest pT
m’s
>90% p/K decay
Algorithm: at least
one cluster/m >threshold
and
“Global”<threshold or
Di-muon>threshold
Nominal
threshold
Nominal threshold
Niko Neufeld
LHCb Trigger, RICH-2004

30. Dedicated HLT RICH algorithms
First step: parameterise ring distortions so that problem
can be solved on HPD planes, without having to solve
quartic equation for each pixel-track combination:
RICH 2
Study distortion by plotting local radius - mean radius
vs photon azimuth. Works well – but not so good for aerogel
Niko Neufeld
LHCb Trigger, RICH-2004

31. Local HLT algorithms: principle
Plot Cherenkov angle for nearby hits to track of interest and
look for peak – below shown integrated over many tracks
Aerogel
Gas, R2
Gas R1
Method OK for gas rings, but is not so good (yet) for aerogel
Niko Neufeld
LHCb Trigger, RICH-2004

32. Selecting the kaon band :
Cutting the pion band :
Selecting the kaon band :
RICH 1
RICH 1
Cut on (rS/B max-rtrue pion)
Cut on (rS/B max-rtrue kaon)
Niko Neufeld
LHCb Trigger, RICH-2004