Performance of the ATLAS Trigger System in 2015

ATLAS Collaboration

Source-linked AI summary

Performance of the ATLAS Trigger System in 2015

ATLAS Collaboration

arXiv:1611.09661v2Nov 29, 2016hep-ex

TL;DR

Higher Run 2 energy, luminosity, and pile-up required ATLAS to upgrade its trigger and data-acquisition system to preserve physics-driven selection rates. The paper reviews those changes and evaluates 2015 trigger performance, finding successful commissioning and high efficiencies, while documenting an early high-energy calorimeter-timing issue that was fixed in firmware for 2016.

Problem
Higher centre-of-mass energy, luminosity, and pile-up threatened trigger-rate limits at the thresholds needed for the ATLAS physics programme.
Method
The paper reviews LS1 trigger and DAQ upgrades and measures trigger-signature performance and efficiencies with 2015 13 TeV collision data.
Results
The ATLAS trigger system was successfully commissioned with 2015 13 TeV data, with first performance studies of trigger signatures and offline-referenced efficiencies presented.
Takeaways & Limitations
The redesigned trigger strategy maintained unprescaled single-electron and single-muon thresholds around 25 GeV to collect most leptonic W and Z decays.
Takeaways & Limitations
A filtering issue caused early trigger signals and incomplete events for a small fraction of very high-energy events until firmware correction was deployed in 2016.

Abstract

from arXiv · show

During 2015 the ATLAS experiment recorded 3.8 fb$^{-1}$ of proton--proton collision data at a centre-of-mass energy of 13 TeV. The ATLAS trigger system is a crucial component of the experiment, responsible for selecting events of interest at a recording rate of approximately 1 kHz from up to 40 MHz of collisions. This paper presents a short overview of the changes to the trigger and data acquisition systems during the first long shutdown of the LHC and shows the performance of the trigger system and its components based on the 2015 proton--proton collision data.

1 Introduction

The ATLAS trigger selects events for later study from LHC bunch crossings. Run 2 conditions increased rates, motivating extensive upgrades during LS1, whose scope and performance are reviewed using 2015 data.

The trigger decides whether events from individual bunch-crossing interactions are retained for later study.
13 TeV collisions, higher luminosity, and increased pile-up would have exceeded allowed trigger rates without upgrades at Run 2 thresholds.
The first long shutdown was used to improve nearly every trigger-system component before Run 2.
The paper summarizes LS1 changes and evaluates trigger signatures using 25 ns, 2015 proton–proton collision data alongside Monte Carlo simulation.

2 ATLAS detector

ATLAS is a general-purpose detector with near-complete solid-angle coverage, combining inner tracking, calorimetry, and muon detection. Its trigger and data-acquisition system uses hardware-based L1 and software-based HLT components.

ATLAS provides almost full solid-angle coverage around the interaction point with forward-backward symmetry.
The detector comprises an inner detector, calorimeter system, and muon spectrometer, with the inner detector inside a 2 T solenoid.
The inner detector reconstructs tracks within |η| < 2.5 using Pixel, SCT, and TRT subsystems, with the IBL added during LS1.
The TDAQ system combines a hardware-based Level-1 trigger with a software-based high-level trigger.
The Central Trigger Processor forms the L1 decision from calorimeter, muon, and other subsystem inputs while applying preventive dead-time.

3 Changes to the Trigger/DAQ system for Run 2

Run 2 trigger and DAQ upgrades addressed higher rates from 13 TeV operation, shorter bunch spacing, and pile-up. The changes included redesigned central architecture, upgraded calorimeter and muon logic, and expanded trigger selectivity, while L1Topo and FTK were not used for the 2015 results.

Trigger and DAQ architecture: L1Topo and FTK were commissioned during 2015 and were not used for the results presented.
Run 2 operating conditions: 13 TeV operation increased trigger rates by 2.0 to 2.5 times at the same luminosity and criteria, while 25 ns spacing added out-of-time pile-up.
Trigger and DAQ architecture: The HLT merged the separate Run 1 L2 and Event Filter farms into one homogeneous farm for resource sharing and simpler hardware and software.
Level-1 calorimeter trigger: Dynamic, bunch-by-bunch pedestal subtraction compensated pile-up effects and linearised the L1 missing-transverse-momentum rate versus instantaneous luminosity.
Level-1 calorimeter trigger: The upgraded calorimeter merger transmitted trigger-object locations and energies to L1Topo, with above 2 Tbps total output bandwidth and increased selectivity.
Level-1 muon trigger: A TGC-FI coincidence in 1.3 < |η| < 1.9 reduced the end-cap muon trigger rate by up to 60% while losing about 2% of offline reconstructed muons.

4 Trigger menu

The 2015 trigger menu organized selections by purpose and physics signature while optimizing thresholds, rates, bandwidth, and computing constraints. It maintained broad physics coverage, supported commissioning and calibration, and used dedicated streaming and processing strategies for efficient data handling.

Trigger menu composition: The trigger menu comprised primary, support, alternative, backup, and calibration triggers with distinct physics, monitoring, reconstruction, rate, or calibration roles.Primary triggers covered electrons, photons, muons, taus, jets, missing transverse momentum, and low-pT dimuon B-physics signatures.
Rate and bandwidth constraints: The menu was optimized across luminosity ranges to satisfy the 100 kHz L1 and 1,000 Hz average HLT physics-output limits.These constraints reflected detector readout and offline computing capabilities.
Trigger menu strategy: Single-electron and single-muon thresholds were kept around 25 GeV to collect most leptonic W and Z decays, with a simpler and more robust strategy than many analysis-specific triggers.Dedicated multi-object triggers supplemented this strategy.
Trigger rates: At a peak luminosity of 4.5 × 10^33 cm−2 s−1, single-electron and single-muon triggers formed a large fraction of the total rate, while support triggers contributed about 20%.B-physics triggers received a large bandwidth share at these relatively low luminosities, and backup triggers added rate because 2015 commissioning time was limited.
Event streaming: HLT streams separated physics, prompt calibration and data-quality, debug, partial-event-building, and specialized overlay data according to their downstream use.The single Main physics stream replaced three Run 1 physics streams and reduced duplication-related storage and reconstruction CPU needs by roughly 10%.
HLT processing time: At the highest 2015 luminosity point, mean HLT processing time was approximately 235 ms, with an 80 kHz L1 rate using 67% of 28,000 available CPU cores.Processing time was distributed mainly across inner-detector tracking, muon-spectrometer reconstruction, and calorimeter reconstruction, and could not be naively scaled because the menu changed with luminosity.
Dedicated menus: Special menus addressed low-bunch operation, enhanced minimum-bias samples, and beam-separation scans, including protection against detector-wire-bond resonances.For low bunch counts, the IBL and SCT required avoidance of dangerous trigger frequencies.

5 High-level trigger reconstruction

The Run 2 high-level trigger uses staged, region-of-interest reconstruction with redesigned tracking and calorimeter algorithms. Performance remained high for electron, muon, tau, and b-jet signatures while staged processing reduced computational demands.

High-level trigger architecture: HLT reconstruction uses fine-granularity calorimeter, muon-spectrometer, and inner-detector information, usually processed in L1-defined regions of interest.Data are retrieved on demand, and most triggers use a two-stage approach to reduce processing time.
Inner-detector tracking: The Run 2 ID trigger was redesigned within the merged HLT and incorporated IBL information, improving tracking and impact-parameter resolution.Provision was also made for later inclusion of FTK tracks.
Inner-detector tracking performance: Electron tracking efficiencies exceed 99% across pseudorapidity for both fast and precision tracking, with a small low-pT loss from bremsstrahlung.The measurement uses offline tracks with pT > 20 GeV and tight offline electron candidates.
Inner-detector tracking performance: Muon tracking efficiency is significantly better than 99%, while impact-parameter resolution is better than 17 µm for fast and 15 µm for precision tracking above 20 GeV.These results are measured relative to loose offline muon candidates with pT > 6 GeV.
Multiple-stage tracking: Two-stage tau tracking exceeds 96% efficiency everywhere in the second stage and reaches better than 99% above 2 GeV, while precision impact-parameter resolution is around 20 µm above 10 GeV.The first-stage efficiency rises from 94% at 2 GeV to better than 99% above 5 GeV because its narrow RoI loses low-pT tracks.
Tracking timing: For tau tracking, two-stage precision processing averages 4.8 ms per RoI versus 12 ms for single-stage processing, owing to the reduced z extent and fewer tracks.Fast processing averages 23 ms and 21 ms for the first and second stages, compared with approximately 66 ms for single-stage tracking.

5.3 Tracking in the muon spectrometer

The muon spectrometer trigger reconstructs candidates in fast and precision stages, combining muon-system and inner-detector information. Online track parameters agree well with offline results, while the precision stage requires staged execution because of its processing time.

Muon-trigger reconstruction: L1 identifies muons from RPC or TGC hit coincidences, estimates pT using six thresholds, and passes RoI information to the HLT for refinement.The HLT adds precision MDT and CSC measurements.
Fast reconstruction: Fast reconstruction fits MDT data and assigns pT to create MS-only candidates, then back-extrapolates their tracks toward the interaction point.The fast stage uses trigger-specific processing within the L1-defined RoI.
Precision reconstruction: Precision reconstruction starts from fast-stage RoIs, builds segments and tracks, and combines MS candidates with ID tracks to form combined muons.MS-only candidates remain available when no matching ID track exists or for specialised triggers.
Resolution: Combined-muon momentum resolution is better than MS-only resolution, especially at low pT, while barrel resolution is generally better than end-cap resolution.The advantage comes from the higher precision of ID measurements, and detector granularity contributes to the barrel/end-cap difference.
Resolution: Online and offline muon track parameters show good agreement, with η and φ residual widths decreasing as pT increases.The comparison uses residual distributions in pT bins and Gaussian fits to their cores.
Processing time: Precision muon reconstruction uses too much time to run alone at the full L1 rate, motivating the fast-plus-precision two-stage design.Processing times per RoI are compared for fast MS-only, fast combined, and precision algorithms.

6 Trigger signature performance

The trigger-signature studies define object and event selections and evaluate their efficiencies using offline-selected data. Tag-and-probe and bootstrap methods provide unbiased efficiency measurements, with simulation used for corrections where applicable.

Performance methodology: Trigger signatures apply selection criteria to reconstructed leptons, hadrons, and global quantities such as missing transverse momentum.The paper lists primary 2015 triggers together with output rates and performance.
Efficiency measurements: The tag-and-probe method measures efficiency using paired offline objects, with one object triggering the event and the other serving as the probe.An example is a pair of electrons from a Z → ee decay.
Efficiency measurements: The bootstrap method determines the efficiency of a higher-threshold trigger from events selected by a lower-threshold trigger.
Efficiency measurements: Efficiencies are computed relative to offline-selected data, and the measured-to-simulated efficiency ratio is used as a correction factor in physics analyses.Unless otherwise specified, studies use 3.2 fb^-1 of 25 ns data collected during 2015.

6.1 Minimum-bias and forward triggers

The minimum-bias and high-multiplicity triggers recorded inelastic and high-track-count events at 13 TeV using MBTS and track-based selections. MBTS_1 and MBTS_2 rapidly reach full efficiency with multiplicity, whereas the two-sided MBTS_1_1 trigger requires roughly 15 tracks because of diffractive events.

Minimum-bias triggers: The MBTS minimum-bias trigger is highly efficient for inelastic events containing only two charged particles with pT > 100 MeV and |η| < 2.5.Minimum-bias triggers support measurements of total cross sections, hadronisation, diffraction, and other non-perturbative properties.
Minimum-bias triggers: More than 200 million interactions were recorded in one week, while the loosest L1_MBTS_1 trigger operated at an average rate of 1.0–1.5 kHz.The data were collected at √s = 13 TeV with a maximum interaction rate of about 15 kHz.
High-multiplicity triggers: The high-multiplicity trigger requires at least 900 SCT space-points and 60 tracks with pT > 400 MeV, reducing HLT track-finding complexity to an acceptable level.
Trigger efficiencies: L1_MBTS_1 and L1_MBTS_2 rise from about 95% efficiency in the first multiplicity bin to 100%, while L1_MBTS_1_1 approaches 100% only around 15 tracks.The two-sided requirement of L1_MBTS_1_1 lowers efficiency at low multiplicity because diffractive events often produce particles on only one detector side.
High-multiplicity triggers: High-multiplicity trigger efficiency is evaluated versus the number of offline tracks associated with the primary vertex for multiple space-point, track, and pT selections.The L1_MBTS_1 trigger provides the control sample for these turn-on curves.

6.2 Electrons and photons

Electron and photon triggers combine fast, staged reconstruction with offline-like precision algorithms to select physics signatures across a broad energy range. In 2015, their rates scaled linearly with instantaneous luminosity, while photon-trigger efficiencies agreed very well between data and simulation.

Trigger menu: Electron and photon triggers cover energies from a few GeV to several TeV for precision measurements and searches.Low-ET triggers support J/ψ →ee, diphoton, and low-mass Drell–Yan measurements.
Reconstruction and selection: HLT reconstruction proceeds from fast calorimeter clustering and tracking to precise offline-like algorithms at reduced rates.The staged design enables more CPU-intensive algorithms later in the trigger sequence.
Reconstruction and selection: Electron identification uses likelihood operating points, with trigger variables based on calorimeter shower shapes, track quality, matching, and TRT particle identification.The trigger uses loose, medium, tight, and very loose likelihood working points.
Trigger menu: 24 GeV is the transverse-energy threshold of the lowest-threshold unprescaled single-electron trigger, seeded by L1_EM20VH at 20 GeV.The trigger requires medium likelihood identification and applies an ET-dependent hadronic veto.
Performance: Electron and photon trigger rates scale linearly with instantaneous luminosity.Figures 24 and 25 show the L1 and HLT rates for selected single- and multi-object triggers.
Performance: Very good agreement is observed between photon-trigger efficiencies measured in data and those obtained from Monte Carlo simulation.The comparison covers the main single-photon and diphoton triggers as functions of offline photon ET and pseudorapidity.

6.3 Muons

Muon triggers use combined reconstruction, isolation, and complementary thresholds to select muon signatures over a wide transverse-momentum range. Their rates scale linearly with luminosity, while HLT efficiency relative to L1 is close to 100% in both barrel and end-cap regions.

Physics coverage: Muon triggers cover transverse momenta from a few GeV to several TeV for precision measurements and searches.Thresholds of 4–10 GeV support J/ψ →µµ, low-pT dimuon, and Z →ττ measurements.
Trigger selection: The lowest-threshold unprescaled single-muon trigger requires 20 GeV combined-muon pT and loose isolation, seeded by L1_MU15.Isolation reduces the rate by approximately 2.5 with negligible efficiency loss; mu50 complements it above 50 GeV.
Rates: Dimuon trigger rates scale linearly with instantaneous luminosity.The lowest-threshold unprescaled dimuon trigger, 2mu10, requires two combined muons above 10 GeV.
Efficiency: The L1_MU15 efficiency is close to 70% in the barrel and 90% in the end-caps.The difference reflects geometrical acceptance and local detector inefficiencies.
Efficiency: HLT efficiency relative to L1 is close to 100% in both the barrel and end-caps.The efficiency is shown versus probe-muon transverse momentum and azimuthal angle.

6.4 Jets

The 2015 jet trigger system supports single-jet, multi-jet, HT, large-R, and analysis-specific selections through flexible HLT reconstruction and calibration. Data and simulation show good efficiency agreement, while Trigger-Level Analysis increases the available low-pT jet sample by recording partial events.

Jet reconstruction: HLT jet reconstruction supports anti-kt jets with R = 0.4 or R = 1.0 and multiple input and calibration configurations.Options include RoI-restricted clusters, local calibration, pile-up subtraction, response corrections, and reclustering.
Trigger menu: The jet menu includes single-jet, multi-jet, HT, and analysis-specific triggers with thresholds set on jet ET, multiplicity, or summed HT.HT is the scalar sum of transverse energies of jets passing the relevant selection.
Trigger rates: 18 Hz is the rate of the lowest-threshold unprescaled standard single-jet trigger, j360, at 5 × 10^33 cm^-2 s^-1.The same passage reports rates of 6, 20, 15, and 12 Hz for 3j175, 4j85, 5j60, and 6j45, respectively.
Performance: Good agreement is observed between data and simulation for single-jet efficiencies in central and forward calorimeter regions.The sharp HLT turn-ons result from good agreement between trigger and offline jet energy scales.
Performance: Multi-jet efficiencies are dominated by the Nth-leading jet and agree across multiplicities with common thresholds and between HLT data and simulation.The comparison is made as a function of the Nth-leading jet transverse momentum.
Trigger-Level Analysis: Trigger-Level Analysis increases event recording rates by a factor of 100 and significantly increases the number of low-pT jets.TLA records only relevant HLT jet objects in a dedicated stream, addressing bandwidth limits for sub-TeV dijet-resonance searches.

6.5 Tau leptons

Tau triggers combine calorimeter, tracking, and multivariate identification to select hadronic tau decays despite similar high-rate jet backgrounds. Their performance is measured with Z →τµτhad tag-and-probe events, with efficiencies evaluated after L1 and HLT selections.

Motivation: Hadronic tau triggers are crucial because about 65% of tau leptons decay hadronically.Tau-pair final states support Higgs coupling measurements and searches for heavier Higgs bosons or other resonances.
Motivation: Tau triggering is challenging because jets are produced copiously and can have features similar to hadronic tau decays.The algorithms exploit narrow calorimeter deposits and small associated track multiplicities.
Reconstruction and selection: HLT tau selection applies an energy requirement, track-multiplicity selection, precision tracking, and a BDT-based identification score.The online BDT uses variables closely following offline counterparts and shares the same training as offline identification.
Trigger menu: All tau triggers apply L1 isolation, HLT track multiplicity, and online medium-identification requirements.The menu includes single-tau and combined τ + X triggers, where X may be an electron, muon, second tau, or missing transverse momentum.
Efficiency measurement: Tau-trigger efficiency is measured with tag-and-probe events containing an offline tau above 25 GeV, selected using a single-muon trigger.Background contributions from W+jets and multijet events are estimated in data and subtracted.
Performance: The tau25_medium trigger requires an isolated L1 RoI above 12 GeV and an HLT tau above 25 GeV passing track-multiplicity and medium-identification criteria.Its candidate distributions are shown for transverse momentum, pseudorapidity, track multiplicity, and BDT score.

6.6 Missing transverse momentum

The 2015 missing-transverse-momentum trigger used several calorimeter-based algorithms, including pile-up corrections, to control rates while retaining physics efficiency. L1 and HLT performance was evaluated through rates, efficiencies, linearity, and resolution.

The missing-transverse-momentum trigger supports searches with jets, energetic muons, and mult-object final states, while also providing samples for detector studies.
Performance: The L1 improvements enabled the L1_XE50 trigger to operate without prescale throughout 2015.
HLT reconstruction: Five HLT algorithms based on cells, jets, or topo-clusters were commissioned, including two methods designed to suppress pile-up effects.The algorithms shared time-consuming clustering, so running them in parallel required only a small additional CPU cost.
Pile-up correction: Pile-up suppression estimated energy contributions in η rings or calorimeter towers before recalculating missing transverse momentum.The ring method omitted topo-clusters above 2σ, while the tower-based method fitted pile-up contributions subject to an event-wide constraint.
Trigger rates: 700 Hz and 50 Hz were the approximate output rates for the unprescaled L1_XE50 and 70 GeV HLT xe triggers, respectively, at 5 × 10^33 cm^-2 s^-1.
Performance: 95–99% was the reported efficiency range, depending on event selection, with topo-cluster algorithms showing higher efficiency in the turn-on region than the cell-based algorithm.

6.7 b-Jets

The b-jet trigger identifies heavy-flavour jets using tracks, vertices, and multivariate tagging, enabling lower jet thresholds for targeted physics signatures. Run 2 improvements included upgraded tracking and adaptation of offline MV2 tagging algorithms.

b-jet triggers identify heavy-flavour content online for fully hadronic final states relevant to Higgs, top-quark, and beyond-the-Standard-Model searches.
b-jet identification: The trigger exploits displaced tracks, secondary vertices, and b-hadron decay properties to distinguish b-jets from light-quark jets.
Run 2 improvements: The IBL improved track impact-parameter resolution, benefiting b-jet identification and trigger performance.
Run 2 improvements: MV2c20 combines IP3D, SV1, and JetFitter inputs; it was commissioned in 2015 and became the baseline b-tagging algorithm for 2016.
Performance: A significant gain in trigger efficiency was observed when moving from the earlier b-tagging algorithms to MV2 algorithms.
Trigger thresholds: A loose b-tagging requirement lowered the unprescaled single-jet threshold from 360 GeV to 225 GeV.

6.8 B-physics

B-physics triggers primarily select dimuon signatures using staged L1 and HLT requirements on muon thresholds, invariant mass, charge, vertices, and additional tracks. Their configurations balance low transverse-momentum reach against trigger-rate constraints.

B-physics triggers target b-hadron and prompt quarkonium decays containing dimuon final states, with additional selections for semileptonic decay hypotheses.
Trigger strategy: Two-muon L1 triggers substantially reduce rates relative to single-muon triggers but lose efficiency at high transverse momentum when the muons become closely separated.
Trigger strategy: HLT selections used invariant-mass windows for J/ψ, B, and Υ dimuon signatures, alongside opposite-sign and common-vertex requirements.
Rates and prescales: The L1_2MU4-seeded HLT triggers were unprescaled up to 4 × 10^33 cm^-2 s^-1; above that luminosity, higher-threshold seeds were used, with a 15% overall event loss from the former.
Efficiency measurement: The efficiency of the opposite-sign and vertex-quality selection was measured using supporting dimuon triggers and offline-selected J/ψ candidates.

7 Conclusion

ATLAS substantially upgraded its trigger and data-acquisition systems during LS1 for Run 2, then commissioned and evaluated them with 13 TeV collision data collected in 2015.

The LS1 programme introduced numerous trigger upgrades and new systems across the L1 and HLT architectures.
L1 upgrades: L1 improvements included dynamic pedestal correction, new muon coincidence logic and chambers, a larger threshold capacity, a 100 kHz output limit, and a new topological processor.
HLT upgrades: The HLT architecture unified the Run 1 Level-2 and Event Filter scheme, while upgraded software moved algorithms closer to offline reconstruction and used the IBL.
2015 commissioning: The trigger menu was redesigned for higher rates and luminosity, and the system was successfully commissioned with 13 TeV, 25 ns data collected during 2015.