The upgrade of the LHCb trigger for Run III

Size: px

Start display at page:

Download "The upgrade of the LHCb trigger for Run III"

Beatrice Ball
5 years ago
Views:

1 The upgrade of the LHCb trigger for Run III CERN The LHCb upgrade will take place in preparation for data taking in LHC Run III. An important aspect of this is the replacement of the hardware trigger by implementing a full software trigger system. The progress and plans towards this objective are presented, including studies relating to the tracking and reconstruction sequence and the trigger output bandwidth division. PoS(EPSHEP7)8 The European Physical Society Conference on High Energy Physics July, 7 Venice Speaker. On behalf of the LHCb collaboration. c Copyright owned by the author(s) under the terms of the Creative Commons AttributionNonCommercialNoDerivatives. International License (CC BYNCND.).

To cope with this change in luminosity a new trigger paradigm will be adopted, with the current twostage hardware plus software trigger being replaced by a fully software trigger.

2 . Introduction The LHCb experiment will be upgraded in preparation for data taking during LHC Run III. The instantaneous luminosity will increase by a factor of five from L = cm s in Run I and Run II to L = cm s in Run III. To cope with this change in luminosity a new trigger paradigm will be adopted, with the current twostage hardware plus software trigger being replaced by a fully software trigger. Figure shows a comparison of the trigger strategies for (left) Run II and (right) Run III. Removal of the hardware trigger is expected to facilitate an increase in the Figure : The LHCb trigger strategy in (left) Run II and (right) Run III. trigger of purely hadronic decays modes by around a factor of two. It is also important to note that the challenges faced by the trigger system will also change, the rate of beauty and charm production mean that it is not sufficient to simply separate signallike and backgroundlike decay topologies. About % (%) of events will contain a reconstructible charm (beauty) hadron. Exclusive selections will be the standard, with some inclusive triggers remaining to preserve the breadth of the physics programme. To provide high purity samples with high these exclusive selections should be close to the final offline selection used in current analyses. Care must be taken with such advanced selections because the trigger will become more sensitive to detector performance effects, such as detection asymmetries, so the realtime alignment and calibration procedures, developed for LHC Run II, will become even more valuable. PoS(EPSHEP7)8. Tracking and reconstruction sequence For full details of the tracking and reconstruction sequences please see Ref. []. The strategy builds on that of LHCb in Run II (see Fig. ), with the fast stage providing the reconstruction for the first part of the selection process, where the accepted candidates are buffered to disk. Then the realtime alignment and calibration tasks are executed before the best stage provides the full

3 reconstruction. It is important to note that no further processing will be performed, the online best reconstruction will be of offline quality. The performance of the fast stage is crucial for the upgrade trigger project. A first tuning of the algorithms, based on simulations of the expected Run III conditions, to optimise both speed and physics performance has been made. The environment in the upgrade era at LHCb is challenging; the average number of primary vertices per event is a factor of two or three times larger. Despite this, preliminary studies of the primary vertex resolution, shown in Fig., look promising. In fact, the resolution in both (left) x and (right) z directions is better than for Run II. The fit function is given by σ(n) = A C, (.) NB where A, B and C and free parameters and N is the number of tracks associated to the vertex. Resolution [µm] LHCb Preliminary Unofficial A [µm] 78. ±. Upgrade: B.6 ±. C [µm]. ±. A [µm] 9.8 ± 8. Run II: B.86 ±. C [µm]. ±. Upgrade Run II N Resolution [µm] LHCb Preliminary Unofficial A [µm] 96. ±. Upgrade: B.8 ±. C [µm].7 ±. A [µm] 7. ± 9.6 Run II: B. ±. C [µm] 8. ±. Upgrade RunII Figure : Primary vertex resolution in the (left) x and (right) z directions for the fast stage as a function of the number of tracks associated to the vertex. The fit function and fitted parameters are as described in the text. Figure reproduced from Ref. []. Studies of fake track (ghost) rejection, tracks reconstructed from pseudorandom combinations of hits, are also underway. Figure shows the performance of a multivariate classifier designed to reject fake tracks applied to long tracks (tracks with hits in atleast the vertex locator and tracking stations downstream of the magent). Optimisation is in progress, but the latest training (red) is close to performing at the same level seen during Run II. The timing and performance of the fast stage is in line with that expected from the upgrade trigger design report []. The timing per event is currently around 6ms, but this is expected to improve in the future. Tracking efficiencies at the fast stage are slightly better than expected, again with more improvements to come. Throughput performance targets are challenging to meet because the hardware performance growth at equal cost is slowing. Therefore a lot of effort is being spent to design and write new software that fully exploits the multiprocessor paradigm. A new computing technical design report is expected early next year. For more details on the timing and performances studies please see Ref. []. N PoS(EPSHEP7)8

4 mance update Ref: LHCbPUB7 Issue: Date: February, 7 LHCb trigger upgrade VELO Tracks.8.6 Long Tracks LHCb Preliminary Track χ/dof.8 Track χ/dof. Figure : Comparison of fake track (ghost) rejection and signal, with (inset) a zoom of the optimal region, obtained with a multivariate classifier. Figure reproduced from Ref. []. Upstream Tracks Downstream Tracks.8. Trigger bandwidth division.6 The division of the trigger. output bandwidth, the data sample saved to offline storage, is an other challenge for the upgrade trigger. Note that the total bandwidth available is limited by the.8. Default GP available disk space rather than the network infrastructure or the trigger itself. Reduction of the Track χ /dof.8 event size to save more signal events in the same amount of disk space has already been used at. LHCb for highrate channels []. This will become the standard approach in the upgrade era. The proof of principle study documented in Ref. [] uses an automated method to divide the output bandwidth between four charm decay modes. The bandwidth assigned per channel will ultimately be limited by the total number of channels and the physics priority decided by the col.8 laboration. In this study a multivariate classifier is used to tune the output bandwidth consumed by TT Tracks each channel. A genetic algorithm is then used to assign the bandwidth per channel by minimising.6 the following χ function by varying the classifier response for each channel χ =..8 Track χ/dof channel i εi wi max. εi (.) Here wi is the channel.weight for decay mode i as decided by the collaboration (unity for each channel in this study), εi is the signal at a given classifier requirement and εimax is the Signal maximum when it is allowed consume the full available bandwidth. A penalty isons of signal and channel ghost track rejection fromto the most recent ghost probability term is applied when the totaldistribution available bandwidth is exceeded Signal and bandwidth d to the previous version. The track /ndof is overlaid, and an []. insets are included on around 9%. are calculated from simulated samples. An example of the output from the bandwidth division algorithm is shown in Figure. Ultimately the signal efficiencies achieved will depend on the analysts, and theirability to define powerful selections using machine learning techniques. ficiency is determined on the Bs! signal sample. The is determined as reconstructed candidates with respect to MC particles that have left sufficient clusters. Summary ctors to be considered reconstructible. The includes the ghost probability n table it can be seen that the for all long tracks is considerably higher in Preparations for the upgrade of the LHCb trigger for Run III are well underway. The predue to the looser momentum requirement. However, high momentum particles from dicted performance of the tracking and reconstruction looks very promising on simulated data. The s are reconstructed in the fast stage with an very similar to what is possible throughput performance will improve with further optimisation coming from significant work in tly more computing time in the best stage. In all cases the performance is equal to or of the trigger TDR with the fast stage having about half the ghost rate: Improvements spillover, larger gaps and ms counteract the more realistic detector description including tance in the SciFi. For the best stage several efficiencies are provided as a function of ability requirement. This will be tuned to meet physics performance and throughput PoS(EPSHEP7)8.

5 Upgrade trigger: Bandwidth strategy proposal Public Note Selection at HLT Bandwidth [MB/s] Bandwidth [MB/s] to follow. π π K π π S π π π π π π S DK K Kπ K π D K K K π π π π π S π S 6 6 Figure : Example of the output of the bandwidth division studies for four charm decay modes, concerning (left) the signal and (right) the bandwidth usage. The red histogram shows the results when each.9 channel is allowed to use all of the available bandwidth and.8 the blue histogram shows the result of the.7 division. The bandwidth limit is defined as 6MB/s. Figure reproduced from Ref. []. the coming years. First studies of the upgrade trigger bandwidth division have been performed as a. proof of principle, with further studies extending this approach to cover the full physics programme References π Figure : Output of the bandwidth division algorithm assuming a bandwidth limit of 6 MB/s. From top to [] R. Aaij et al., Upgrade trigger: Biannual performance update, LHCbPUB7 (7). bottom are the minimum bias bandwidth per channel, the signal efficiencies and the minimum bias rate. On the left, the default scenario in which D! KS 6 [] R. Aaij et al., LHCb Trigger and Online Upgrade events Technical are saved Design with Report, PersistReco CERNLHCC6 is presented, while on the right, only D! KS signal tracks are saved. Distributions (). in red are those for the initial case in which each signal is allocated % of the available bandwidth; in blue are the final results in which each channel shares the bandwidth. [] R. Aaij et al., Tesla: An application for realtime data analysis in High Energy Physics, Comput. Phys. Commun. 8, (6) K K π K π π S π π PoS(EPSHEP7)8 [] C. Fitzpatrick et al., Upgrade In the initial optimisation step " max trigger: Bandwidth strategy proposal, LHCbPUB76 (7). i is determined foreach channel independently, and it can be seen that the full bandwidth is taken by each channel. In the final step the genetic algorithm determines the MVA response necessary to minimise the, resulting in a distribution of the output bandwidth among the channels. The D! K K channel is considerably less pure than the other channels due to the π lack of a m requirement. As a result, it receives proportionally more K D K π of the available π π bandwidth π S in order to obtain a similar signal. Conversely, the D! KS mode is very pure: At a similar to the other channels it uses the same Figurebandwidth : Output even of thethough bandwidth this bandwidth division algorithm is driven assuming by a the larger event size in the Turbo PersistReco bottom case. are Moving the minimum to Turbo bias only, bandwidth where D per! channel, KS the has signal effi left, the default scenario in which D a similar event size to the other channels, it is considerably more efficient for a! K similar S events are saved rate. right, only D! KS signal tracks are saved. Distributions in re The evolution of the MVA selection signal is is shown allocated in Figure % of the over available the range bandwidth; from in MB/s blueto are the fi GB/s. The for each channel rises bandwidth. rapidly as the bandwidth limit is increased, as expected, and plateaus as " channel s approaches " channel max. If MB/s was available on average per channel, then charm trigger efficiencies of % could be expected in the upgrade. Allowing for MB/s would increase this to %, assuming the MVA In performance the initial optimisation is indicative. step " max i is determined for each ch that the full bandwidth is taken by each channel. In the final s MVA response necessary to minimise the, resulting in a dist the channels. The D! K K channel is considerably les lack of a m requirement. As a result, it receives proportiona page 7 order to obtain a similar signal. Conversely, the D! to the other channels it uses the same bandwidth e the larger event size in the Turbo PersistReco case. Moving a similar event size to the other channels, it is considerably m D K D K K π π S The evolution of the MVA selection is shown in Fi π π

The upgrade of the LHCb trigger for Run III

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