The upgrade of the LHCb trigger for Run III Mark Whitehead on behalf of the LHCb collaboration
Introduction LHCb upgrade for Run III Detector upgrades to cope with increased luminosity Run II L =4 32 cm 2 s (3 TeV ) Run III L =2 33 cm 2 s (4 TeV ) Outline Trigger strategy in Run III Reconstruction Bandwidth studies CERNLHCC246 Trigger TDR LHCbPUB275 LHCbPUB276 8th July 27 EPS 27 Venice M. Whitehead 2
Trigger in Run II and III Trigger strategy changes from Run II to Run III Remove hardware trigger Increased output rate to storage 8th July 27 EPS 27 Venice M. Whitehead 3
A paradigm shift from Run II Trigger in Run III Rates of beauty and charm so high in the upgrade regime that the trigger will not just have to separate signal and background decay topologies Effectively separating signal decays from other signal decays 24% of events will contain a reconstructible charm hadron 2% will contain a beauty hadron Select specific signal channels while suppressing others Exclusive selections will be the standard Retain some inclusive triggers for bredth of the physics programme Should be almost the offline selections aim for high purity and efficiency More sensitivity to detector performance effects (e.g. asymmetries) Realtime alignment and calibration will be crucial 8th July 27 EPS 27 Venice M. Whitehead 4
Tracking and reconstruction sequence For details on the Run II realtime tracking and alignment performance and developments please see Agnieszka Dziurda s talk from Thursday 8th July 27 EPS 27 Venice M. Whitehead 5
Tracking and reconstruction Take advantage of the Run II trigger strategy Perform a fast reconstruction for realtime alignment and calibration Second, best, stage performs the full the reconstruction Online quality = offline quality no need for further processing Figure : A schematic view of the fast tracking stage. LHCbPUB275 8th July 27 EPS 27 Venice M. Whitehead 6
The fast stage Fast stage performance vital for the upgrade trigger Resolution [µm] First tuning for LHC upgrade conditions to optimise both speed and physics performance Challenging environment average number of PVs is 5 times higher Primary vertex resolution looks impressive 45 4 35 3 25 2 5 A [µm] Upgrade: B C [µm] Run II: LHCb 5 Preliminary Unofficial 5 A [µm] B C [µm] 78.5 ± 4..64 ±.3.3 ±.3 59.8 ± 8.3.86 ±.2 2.5 ±.2 x direction Upgrade Run II N Resolution [µm] 3 25 2 5 5 LHCb Preliminary Unofficial 5 A [µm] 926. ± 5.4 Upgrade: B.84 ±.2 C [µm].7 ±.2 A [µm] 75. ± 9.6 Run II: B.4 ±.2 C [µm] 8. ±. z direction Fit function (N) = A N B + C A,B and C free parameters N number of tracks associated to vertex 8th July 27 EPS 27 Venice M. Whitehead 7 Upgrade RunII N 3 LHCbPUB275 [µm] 25
The fast stage Fast stage performance vital for the upgrade trigger First tuning for LHC upgrade conditions to optimise both speed and physics performance Challenging environment average number of PVs is 5 times higher Ghost rejection now approaching Run II performance Ghost rejection.8.6.4.2.8 Long Tracks.8 Retrained GP Default GP Track χ 2 /dof.5 Signal efficiency LHCbPUB275 8th July 27 EPS 27 Venice M. Whitehead 8 ction LHCb Preliminary Uses a multivariate classifier New training for Run III Optimisation ongoing
Timing and performance Timing inline with the trigger TDR CERNLHCC246 Where we expected to be but more improvements to come Tracking efficiencies also look promising Timing [ms] Trigger TDR Fast stage Equal to or better than those in the trigger TDR VELO tracking 2. 2. VELOUT tracking.3.5 Throughput performance targets challenging to meet Forward Hardware performance trackinggrowth at.9 equal cost is slowing 2.3 PV A lot finding of work on a new software.4 framework underway. Fully exploit the multiprocessor paradigm total 5.6 6. Computing TDR expected early next year LHCbPUB275 8th July 27 EPS 27 Venice M. Whitehead 9
Timing and performance Timing inline with the trigger TDR CERNLHCC246 Where we expected to be but more improvements to come Tracking efficiencies also look promising Equal to or better than those in the trigger TDR Throughput performance targets challenging to meet Trigger TDR Fast stage Best Stage Ghost probability <.9 <.75 <.3 <. Mostly from the hardware point of view Ghost rate.9% 5.6% 8.8% 5.2% 7.8% 4.2% long A lot of work on a new 42.7% software framework 42.9% underway 9.% 9.8% 88.2% 84.3% long, from Fully B exploit the multiprocessor 72.5% paradigm 72.7% 94.8% 94.6% 93.% 9.6% long, from B, p TDR T >.5GeV 92.3% 92.5% 96.5% 96.4% 95.4% 93.6% expected early next year LHCbPUB275 8th July 27 EPS 27 Venice M. Whitehead
Timing and performance Timing inline with the trigger TDR CERNLHCC246 Where we expected to be but more improvements to come Tracking efficiencies also look promising Equal to or better than those in the trigger TDR Throughput performance targets challenging to meet Hardware performance growth at equal cost is slowing A lot of work on new software underway Fully exploit the multiprocessor paradigm Computing TDR expected early next year LHCbPUB275 8th July 27 EPS 27 Venice M. Whitehead
Output bandwidth division studies 8th July 27 EPS 27 Venice M. Whitehead 2
Output bandwidth division How do we divide up the trigger output bandwidth? This is the output to offline storage Finite disk space limits the output BW not the network or trigger TURBO stream see Giulio Gazzoni's from Thursday Reduced event size more signal events for the same amount of disk space Use an automated method to divide between channels BW per channel defined by number of channels and physics priority Need a way to tune the output BW consumed per channel Here we study it using an MVA classifier response Proof of principle study using 4 charm decays modes Channel D +! K + K + D! K + K D! K + K + D! KS + Event size 4 kb 2 kb 4 kb 7 kb (4kB) LHCbPUB276 8th July 27 EPS 27 Venice M. Whitehead 3
Output bandwidth division Genetic algorithm approach 2 Minimise the by varying the MVA response for each decay w i channel weight ( =. here) " i channel efficiency channels X " max maximum channel efficiency 2 =! i i ( when given the full output BW i LHCbPUB276 2 " i " max i Assign Maximum these channels possible a 6MB/s bandwidth limit between them and Result use of division the algorithm to divide it up Efficiency calculated from signal MC samples Bandwidth calculated from minimum bias MC sample BW[GB/s] = retention rate event size[kb/evt.] 8th July 27 EPS 27 Venice M. Whitehead 4
Output bandwidth division Genetic algorithm approach 2 Minimise the by varying the MVA response for each decay w i channel weight ( =. here) " i channel efficiency channels Upgrade trigger: Bandwidth strategy X maximum channel efficiency 2 proposal Public Note " max =! i i 4 Selection at HLT2 when given the full output BW i LHCbPUB276 2 " i " max i Efficiency.9.8.7.6.5.4 Maximum possible Result of division Efficiency Bandwidth [MB/s].9.8.7.6.5.4 6 5 4 3.3.3 2.2. + + π + + + π + π D K π + π S + D K + + 8th July 27 EPS 27 Venice M. Whitehead 5 z] 6 z] cy.2. 6 + K π + π + + D K + K D K + + K π + π + π π D K D K π + π S π + S π
Output bandwidth division Genetic algorithm approach 2 Minimise the by varying the MVA response for each decay w i channel weight ( =. here) " i channel efficiency channels Upgrade trigger: Bandwidth strategy X maximum channel efficiency 2 proposal Public Note " max =! i i 4 Selection at HLT2 when given the full output BW i LHCbPUB276 2 " i " max i Efficiency Maximum possible.9.9 6.8.8 Signal Result efficiencies of divisionwill ultimately depend on analysts ability to 5.7.7 define.6 powerful selections.6 4 Efficiency Bandwidth [MB/s] Use machine learning in the trigger.5.5 3.4.4 Reduction of the event size, more signal for the same BW usage.3.3 2.2. + + π + + + π + π D K π + π S + D K + + 8th July 27 EPS 27 Venice M. Whitehead 6 z] 6 z] cy.2. 6 + K π + π + + D K + K D K + + K π + π + π π D K D K π + π S π + S π
Summary LHCb upgrade trigger studies well underway Tracking and reconstruction Very promising performance on simulated data Throughput will improve with further optimisation Significant work will be done in the coming years Trigger bandwidth division First look at charm and proof of principle Next step extend studies to full LHCb physics programme Lots more to come in the next couple of years 8th July 27 EPS 27 Venice M. Whitehead 7
Backups 8th July 27 EPS 27 Venice M. Whitehead 8
Upgrade detector 8th July 27 EPS 27 Venice M. Whitehead 9
Best stage Best stage Figure 2: A schematic view of the best tracking stage. LHCbPUB275 8th July 27 EPS 27 Venice M. Whitehead 2
Tracking and reconstruction Timing and efficiency performance (with captions) Table 2: Timing of the fast stage compared to the one of the trigger TDR with GEC cuts applied. Both timing tests are performed using an X565 EFF node described in the text. Timing [ms] Trigger TDR Fast stage VELO tracking 2. 2. VELOUT tracking.3.5 Forward tracking.9 2.3 PV finding.4. total 5.6 6. Table : Track reconstruction efficiencies and ghost rates of the fast and best stages as compared to the trigger TDR. The best stage efficiencies and ghost rates are shown for several values of the ghost probability requirement. Trigger TDR Fast stage Best Stage Ghost probability <.9 <.75 <.3 <. Ghost rate.9% 5.6% 8.8% 5.2% 7.8% 4.2% long 42.7% 42.9% 9.% 9.8% 88.2% 84.3% long, from B 72.5% 72.7% 94.8% 94.6% 93.% 9.6% long, from B, p T >.5GeV 92.3% 92.5% 96.5% 96.4% 95.4% 93.6% LHCbPUB275 8th July 27 EPS 27 Venice M. Whitehead 2
Output bandwidth division Event size is very important change 7kb to 4kb D! K S + from Bandwidth [MB/s] 6 5 4 3 Bandwidth [MB/s] 6 5 4 3 2 2 Efficiency.9.8.7 + + π + + + π + π D K π + π S Efficiency.9.8.7 + + π + + + π + π D K π + π S Huge efficiency gains for zero bandwidth increase.6.6.5.5.4.4.3.3.2.2. D + K K π + + D K + K D + K K π + π S D K π + π. D + K K π + + D K 8th July 6 27 EPS 27 6 Venice M. Whitehead 22 KHz] KHz] + K D + K K π + π S D K π + π LHCbPUB276