High Level Triggering Fred Wickens
2 High Level Triggering (HLT) Introduction to triggering and HLT systems –Why do we Trigger –Why do we use Multi-Level Triggering –Brief description of “typical” 3 level trigger Case study of ATLAS HLT (+ some comparisons with other experiments) Summary
3 Why do we Trigger and why multi-level Over the years experiments have focussed on rarer processes –Need large statistics of these rare events –DAQ system (and off-line analysis capability) under increasing strain limiting useful event statistics Aim of the trigger is to record just the events of interest i.e. Trigger = system which selects the events we wish to study Originally - only read-out the detector if Trigger satisfied –Larger detectors and slow serial read-out => large dead-time –Also increasingly difficult to select the interesting events Introduced: Multi-level triggers and parallel read-out –At each level apply increasingly complex algorithms to obtain better event selection/background rejection These have: –Led to major reduction in Dead-time – which was the major issue –Managed growth in data rates – this remains the major issue
4 Summary of ATLAS Data Flow Rates From detectors> Bytes/sec After Level-1 accept~ Bytes/sec Into event builder~ 10 9 Bytes/sec Onto permanent storage~ 10 8 Bytes/sec ~ Bytes/year
5 The evolution of DAQ systems
6 TDAQ Comparisons
7 Level 1 Time:few microseconds Hardware based –Using fast detectors + fast algorithms –Reduced granularity and precision calorimeter energy sums tracking by masks During Level-1 decision time store event data in front- end electronics –at LHC use pipeline - as collision rate shorter than Level-1 decision time For details of Level-1 see Dave Newbold lecture –2 weeks ago
8 Level 2 Previously - few milliseconds (1-100) –Dedicated microprocessors, adjustable algorithm 3-D, fine grain calorimetry tracking, matching Topology –Different sub-detectors handled in parallel Primitives from each detector may be combined in a global trigger processor or passed to next level few milliseconds (10-100) –Processor farm with Linux PC’s –Partial events received via high-speed network –Specialised algorithms –Each event allocated to a single processor, large farm of processors to handle rate
9 Level 3 millisecs to seconds processor farm –Previously microprocessors/emulators –Now standard server PC’s full or partial event reconstruction –after event building (collection of all data from all detectors) Each event allocated to a single processor, large farm of processors to handle rate
10 Summary of Introduction For many physics analyses, aim is to obtain as high statistics as possible for a given process –We cannot afford to handle or store all of the data a detector can produce! The Trigger –selects the most interesting events from the myriad of events seen I.e. Obtain better use of limited output band-width Throw away less interesting events Keep all of the good events(or as many as possible) –must get it right any good events thrown away are lost for ever! High level Trigger allows: –More complex selection algorithms –Use of all detectors and full granularity full precision data
Case study of the ATLAS HLT system Concentrate on issues relevant for ATLAS (CMS very similar issues), but try to address some more general points
12 Starting points for any HLT system physics programme for the experiment –what are you trying to measure accelerator parameters –what rates and structures detector and trigger performance –what data is available –what trigger resources do we have to use it
13 Interesting events are buried in a sea of soft interactions Higgs production High energy QCD jet production Physics at the LHC B physics top physics
14 The LHC and ATLAS/CMS LHC has –design luminosity cm -2 s -1 (In 2009 from ?) –bunch separation 25 ns (bunch length ~1 ns) This results in – ~ 23 interactions / bunch crossing ~ 80 charged particles (mainly soft pions) / interaction ~2000 charged particles / bunch crossing Total interaction rate10 9 sec -1 –b-physicsfraction ~ sec -1 –t-physicsfraction ~ sec -1 –Higgsfraction ~ sec -1
15 Physics programme Higgs signal extraction important - but very difficult There is lots of other interesting physics –B physics and CP violation –quarks, gluons and QCD –top quarks –SUSY –‘new’ physics Programme will evolve with: luminosity, HLT capacity and understanding of the detector –low luminosity ( ) high PT programme (Higgs etc.) b-physics programme (CP measurements) –high luminosity (2011?) high PT programme (Higgs etc.) searches for new physics
16 Trigger strategy at LHC To avoid being overwhelmed use signatures with small backgrounds –Leptons –High mass resonances –Heavy quarks The trigger selection looks for events with: –Isolated leptons and photons, – -, central- and forward-jets –Events with high E T –Events with missing E T
17 ObjectsPhysics signatures Electron 1e>25, 2e>15 GeVHiggs (SM, MSSM), new gauge bosons, extra dimensions, SUSY, W, top Photon 1γ>60, 2γ>20 GeVHiggs (SM, MSSM), extra dimensions, SUSY Muon 1μ>20, 2μ>10 GeVHiggs (SM, MSSM), new gauge bosons, extra dimensions, SUSY, W, top Jet 1j>360, 3j>150, 4j>100 GeVSUSY, compositeness, resonances Jet >60 + E T miss >60 GeVSUSY, exotics Tau >30 + E T miss >40 GeVExtended Higgs models, SUSY Example Physics signatures
18 ARCHITECTURE 40 MHz TriggerDAQ ~1 PB/s (equivalent) ~ 200 Hz~ 300 MB/sPhysics Three logical levels LVL1 - Fastest: Only Calo and Mu Hardwired LVL2 - Local: LVL1 refinement + track association LVL3 - Full event: “Offline” analysis ~2.5 s ~40 ms ~4 sec. Hierarchical data-flow On-detector electronics: Pipelines Event fragments buffered in parallel Full event in processor farm
19 Selected (inclusive) signatures
20 Trigger design – Level-1 Level-1 –sets the context for the HLT –reduces triggers to ~75 kHz Uses limited detector data –Fast detectors (Calo + Muon) –Reduced granularity Trigger on inclusive signatures muons; em/tau/jet calo clusters; missing and sum E T Hardware trigger –Programmable thresholds –CTP selection based on multiplicities and thresholds
21 Level-1 Selection The Level-1 trigger –an “or” of a large number of inclusive signals –set to match the current physics priorities and beam conditions Precision of cuts at Level-1 is generally limited Adjust the overall Level-1 accept rate (and the relative frequency of different triggers) by –Adjusting thresholds –Pre-scaling (e.g. only accept every 10th trigger of a particular type) higher rate triggers Can be used to include a low rate of calibration events Menu can be changed at the start of run –Pre-scale factors may change during the course of a run
22 Example Level-1 Menu for 2x10^33 Level-1 signatureOutput Rate (Hz) EM25i EM15i4000 MU MU6200 J J J65200 J60 + XE60400 TAU25i + XE MU10 + EM15i100 Others (pre-scaled, exclusive, monitor, calibration)5000 Total~25000
23 Trigger design - Level-2 Level-2 reduce triggers to ~2 kHz –Note CMS does not have a physically separate Level-2 trigger, but the HLT processors include a first stage of Level-2 algorithms Level-2 trigger has a short time budget –ATLAS ~40 milli-sec average Note for Level-1 the time budget is a hard limit for every event, for the High Level Trigger it is the average that matters, so OK for a small fraction of events to take times much longer than this average Full detector data is available, but to minimise resources needed: –Limit the data accessed –Only unpack detector data when it is needed –Use information from Level-1 to guide the process –Analysis proceeds in steps with possibility to reject event after each step –Use custom algorithms
24 Regions of Interest The Level-1 selection is dominated by local signatures (I.e. within Region of Interest - RoI) –Based on coarse granularity data from calo and mu only Typically, there are 1-2 RoI/event ATLAS uses RoI’s to reduce network b/w and processing power required
25 Trigger design - Level-2 - cont’d Processing scheme –extract features from sub-detector data in each RoI –combine features from one RoI into object –combine objects to test event topology Precision of Level-2 cuts –Emphasis is on very fast algorithms with reasonable accuracy Do not include many corrections which may be applied off-line –Calibrations and alignment available for trigger not as precise as ones available for off-line
26 ARCHITECTURE HLTHLT 40 MHz 75 kHz ~2 kHz ~ 200 Hz 40 MHz RoI data = 1-2% ~2 GB/s FE Pipelines 2.5 s LVL1 accept Read-Out Drivers ROD LVL1 2.5 s Calorimeter Trigger Muon Trigger Event Builder EB ~3 GB/s ROS Read-Out Sub-systems Read-Out Buffers ROB 120 GB/sRead-Out Links Calo MuTrCh Other detectors ~ 1 PB/s Event Filter EFP ~ 1 sec EFN ~3 GB/s ~ 300 MB/s TriggerDAQ LVL2 ~ 10 ms L2P L2SV L2N L2P ROIB LVL2 accept RoI requests RoI’s
27 CMS Event Building CMS perform Event Building after Level-1 This simplifies the architecture, but places much higher demand on technology: –Network traffic ~100 GB/s Use Myrinet instead of GbE for the EB network Plan a number of independent slices with barrel shifter to switch to a new slice at each event –Time will tell which philosophy is better
28 t i m e e30i + Signature ecand + Signature e e + e30 + Signature EM20i + Level1 seed Cluster shape Cluster shape STEP 1 Iso– lation Iso– lation STEP 4 pt> 30GeV pt> 30GeV STEP 3 track finding track finding STEP 2 HLT Strategy: Validate step-by-step Check intermediate signatures Reject as early as possible Sequential/modular approach facilitates early rejection LVL1 triggers on two isolated e/m clusters with pT>20GeV (possible signature: Z–>ee) Example for Two electron trigger
29 Trigger design - Event Filter / Level-3 Event Filter reduce triggers to ~200 Hz Event Filter budget ~ 4 sec average Full event detector data is available, but to minimise resources needed: –Only unpack detector data when it is needed –Use information from Level-2 to guide the process –Analysis proceeds in steps with possibility to reject event after each step –Use optimised off-line algorithms
30 Electron slice at the EF TrigCaloRec EF tracking TrigEgammaRec EFTrackHypo Wrapper of CaloRec Wrapper of newTracking Wrapper of EgammaRec EFCaloHypo EFEgammaHypo matches electromagnetic clusters with tracks and builds egamma objects
31 Trigger design - HLT strategy Level 2 –confirm Level 1, some inclusive, some semi- inclusive, some simple topology triggers, vertex reconstruction (e.g. two particle mass cuts to select Zs) Level 3 –confirm Level 2, more refined topology selection, near off-line code
32 Example HLT Menu for 2x10^33 HLT signatureOutput Rate (Hz) e25i40 2e15i<1 gamma60i25 2gamma20i2 mu20i40 2mu1010 j j j11010 j70 + xE7020 tau35i + xE455 2mu6 with vertex, decay-length and mass cuts (J/psi, psi ’, B)10 Others (pre-scaled, exclusive, monitor, calibration)20 Total~200
33 Example B-physics Menu for 10^33 LVL1 : MU6 rate 24kHz (note there are large uncertainties in cross-section) In case of larger rates use MU8 => 1/2xRate 2MU6 LVL2: Run muFast in LVL1 RoI ~ 9kHz Run ID recon. in muFast RoI mu6 (combined muon & ID) ~ 5kHz Run TrigDiMuon seeded by mu6 RoI (or MU6) Make exclusive and semi-inclusive selections using loose cuts –B(mumu), B(mumu)X, J/psi(mumu) Run IDSCAN in Jet RoI, make selection for Ds(PhiPi) EF: Redo muon reconstruction in LVL2 (LVL1) RoI Redo track reconstruction in Jet RoI Selections for B(mumu) B(mumuK*) B(mumuPhi), BsDsPhiPi etc.
34 Matching problem Background Physics channel Off-line On-line
35 Matching problem (cont.) ideally –off-line algorithms select phase space which shrink-wraps the physics channel –trigger algorithms shrink-wrap the off-line selection in practice, this doesn’t happen –need to match the off-line algorithm selection For this reason many trigger studies quote trigger efficiency wrt events which pass off-line selection –BUT off-line can change algorithm, re-process and recalibrate at a later stage So, make sure on-line algorithm selection is well known, controlled and monitored
36 Selection and rejection as selection criteria are tightened –background rejection improves –BUT event selection efficiency decreases
37 Selection and rejection Example of a ATLAS Event Filter (I.e. Level-3) study of the effectiveness of various discriminants used to select 25 GeV electrons from a background of dijets
38 Other issues for the Trigger Efficiency and Monitoring –In general need high trigger efficiency –Also for many analyses need a well known efficiency Monitor efficiency by various means –Overlapping triggers –Pre-scaled samples of triggers in tagging mode (pass-through) –To assist with overall normalisation ATLAS divides each run into periods of a few minutes called a luminosity block. During each block the beam luminosity should be constant and can also exclude any blocks where there is a known problem Final detector calibration and alignment constants not available immediately - keep as up-to-date as possible and allow for the lower precision in the trigger cuts when defining trigger menus and in subsequent analyses Code used in trigger needs to be very robust - low memory leaks, low crash rate, fast
39 Other issues for the Trigger – cont’d Beam conditions and HLT resources will evolve over several years (for both ATLAS and CMS) –In 2009 luminosity low, but also HLT capacity will be < 50% of full system For details of the current ideas on ATLAS Menu evolution see – Gives details of menu for Startup (including single beam running), 10^31, 10^32, 10^33 Description of current menu is very long !! –Trigger Workshop next week in Beatenberg One aim to reduce menus to something more manageable for early running Corresponding information for CMS is at –
40 Summary High-level triggers allow complex selection procedures to be applied as the data is taken –Thus allow large samples of rare events to be recorded The trigger stages - in the ATLAS example –Level 1 uses inclusive signatures (mu’s; em/tau/jet; missing and sum E T ) –Level 2 refines Level 1 selection, adds simple topology triggers, vertex reconstruction, etc –Level 3 refines Level 2 adds more refined topology selection Trigger menus need to be defined, taking into account: –Physics priorities, beam conditions, HLT resources Include items for monitoring trigger efficiency and calibration Try to match trigger cuts to off-line selection Trigger efficiency should be as high as possible and well monitored Must get it right - events thrown away are lost for ever! Triggering closely linked to physics analyses – so enjoy!
41 Additional Foils
42 ATLAS HLT Hardware Each rack of HLT (XPU) processors contains -~30 HLT PC’s (PC’s very similar to Tier-0/1 compute nodes) -2 Gigabit Ethernet Switches -a dedicated Local File Server Final system will contain ~2300 PC’s
43 SDX1|2 nd floor|Rows 3 & 2 CFS nodes UPS for CFS LFS nodes
44 Naming Convention First Level Trigger (LVL1) Signatures in capitals e.g. LVL1 HLTtype EM eelectron gphoton MUmumuon HAtau FJ fj forward jet JEjejet energy JTjtjet TMxemissing energy HLT in lower case: name threshold isolated mu 20 i _ passEF EF in tagging mode name threshold isolated MU 20 I New in : Threshold is cut value applied previously was ~95% effic. point. More details : see :
45 What is a minimum bias event ? - event accepted with the only requirement being activity in the detector with minimal pT threshold [100 MeV] (zero bias events have no requirements) - e.g. Scintillators at L1 + (> 40 SCT S.P. or > 900 Pixel clusters) at L2 - a miminum bias event is most likely to be either: - a low pT (soft) non-diffractive event - a soft single-diffractive event - a soft double diffractive event (some people do not include the diffractive events in the definition !) - it is characterised by: - having no high pT objects : jets; leptons; photons - being isotropic - see low pT tracks at all phi in a tracking detector - see uniform energy deposits in calorimeter as function of rapidity - these events occur in % of collisions. So if any given crossing has two interactions and one of them has been triggered due to a high pT component then the likelihood is that the accompanying event will be a dull minimum bias event.
46 L1 Rates Removing overlaps between single+multi EM gives 18 kHz Total estimated L1 rate with all overlaps removed is ~ 12 kHz Trigger GroupRate (Hz) Multi EM6400 Multi Object5500 Single EM5500 Single Muon1700 Multi Tau470 Single Tau150 Jets80 Multi Muon70 XE50 TOTAL20000
47 L2 Rates Total estimated L2rate with all overlaps removed is 840 Hz Trigger GroupRate (Hz) Electrons310 Muons210* Taus+X180 XE+82 Photons46 B Phys43 Jets22 TOTAL900 * Manually prescaled off pass-through triggers mu4_tile, mu4_mu6 X=anything; + includesJE, TE, anything with MET except taus; Bphys includes Bjet
48 EF Rates Total estimated EF Rate with overlaps removed is 250 Hz Trigger GroupRate (Hz) Muons 80 Electrons67 Tau+X56 B Phys37 Jets25 Photons18 XE+13 Misc13 TOTAL Hz total is in prescaled triggers; 51 Hz of prescaled triggers is unique rate
49 L1 Rates Total estimated L1 rate with all overlaps removed is 46 kHz Trigger GroupRate (Hz) Multi Object30000 Single Muon17000 Multi EM11000 Single EM8100 Multi Tau4300 Single Tau870 Multi Muon690 Jets300 XE300 TOTAL73000
50 L2 Rates Total estimated L2 with all overlaps removed is 1700 (too high!) Trigger GroupRate (Hz) Tau+X820 XE+590 Electrons390 Muons280 3 Objects270 Photons120 B Phys110 Jets33 Misc28 TOTAL2600
51 EF Rates Total estimated EF rates with all overlaps removed is 390 Hz (Fixing L2 will likely come close to fixing EF as well) Trigger GroupRate (Hz) Tau+X187 Electrons77 Muons46 Photons46 BPhys45 3 Objects45 XE+42 Jets11 Misc11 TOTAL510