LHCb L1&DAQ review – April 29, 2003, CERN Massimiliano Ferro-Luzzi 1
Data sizes and running scenarios overview
Data sizes and running scenarios overview
LHCb Trigger and hit multiplicities:
Distributions of VELO and TT clusters for L1
L1 Processing time distribution
Trigger sensitivity to hit multiplicity (cuts)
Motivate various running scenarios for farm design studies
LHCb L1&DAQ review – April 29, 2003, CERN Massimiliano Ferro-Luzzi 2
Current Baseline Scenario
Current Baseline Scenario
L1/DAQ CPU farm designed for following averages:L1 input rate 1.1 MHzL1 output rate 40 KHzL1 CPUs ~ 500HLT CPUs ~ 700L1 processing time ~ 0.5 msHLT processing time ~ 17.5 msL1 uses VELO + TT + L0 objectsHLT uses as L1 + Calo, Muon, Trackers
BUT: what if hit multiplicities are larger than expected ?… larger data sizes, larger processing times.What can we do ?
LHCb L1&DAQ review – April 29, 2003, CERN Massimiliano Ferro-Luzzi 3
Trigger and MultiplicitiesTrigger and MultiplicitiesR
ate
(a.u
)
SPD multiplicity
• Because of inefficiencies (pattern recognition), offline signal content goes down at large multiplicities.
• Events at large multiplicities consume more CPU time and put more data through the network.
• The idea: try to replace those events a.s.a.p. (i.e. at L0) by lower multiplicity events.
• Because of inefficiencies (pattern recognition), offline signal content goes down at large multiplicities.
• Events at large multiplicities consume more CPU time and put more data through the network.
• The idea: try to replace those events a.s.a.p. (i.e. at L0) by lower multiplicity events.
LHCb L1&DAQ review – April 29, 2003, CERN Massimiliano Ferro-Luzzi 4
L1VeloTrackAlg Time distributionsL1VeloTrackAlg Time distributionsBrunel v14r2
tav t no_cut
Mean and RMS in NSPD-bin
LHCb L1&DAQ review – April 29, 2003, CERN Massimiliano Ferro-Luzzi 5
L0 and SPD multiplicity cutL0 and SPD multiplicity cutBrunel v14r2 (fall 2002 production)Bs J/ (e+e-)
1MHz
L0 can cope with a multiplicity cut. What about L1 ?
L0 output fixed at 1MHz
LHCb L1&DAQ review – April 29, 2003, CERN Massimiliano Ferro-Luzzi 6
Sensitivity of L0*L1 to multiplicity: a first lookSensitivity of L0*L1 to multiplicity: a first look
Apply a given cut on SPD multiplicity at L0
Each time, re-adjust L0 thresholds to get always the same L0 minimum bias retention rate
Pass events to L1 and re-adjust threshold on L1 global variable to get always the same L1 minimum bias retention rate
LHCb L1&DAQ review – April 29, 2003, CERN Massimiliano Ferro-Luzzi 7
L1 versus VELO clusters
L1 versus VELO clusters
Small fraction of events with more than (say) 2500 clusters in L0yes MBIA and in offline-selected L0yes signal events
But large fraction of L1 time spent on those MBIA events!
Brunel v14r2
Evt fraction with more than Nvc velo clusters
Time (a.u.) spent on evts with Nvc velo clus.
LHCb L1&DAQ review – April 29, 2003, CERN Massimiliano Ferro-Luzzi 8
Scenarios for farm studies
Scenarios for farm studies
So far, we considered three scenarios (with a fixed # of CPUs):Sc1: Baseline scenario:
L1 processing time scaled such that average is ~0.5 ms. HLT processing time scaled such that average is ~17 ms. L1 output rate = 40 kHz, using only TT+VELO+L0. 500 L1 CPUs + 700 HLT CPUs.
Sc2: Upgrade scenario: L1 processing time scaled such that average is ~1 ms. HLT processing time scaled such that average is ~17 ms. L1 output rate = 12 kHz, using TT+VELO+L0+Trackers ( + Calo sel. crate, muon ). 1000 L1 CPUs + 200 HLT CPUs.
Sc3: Robustness check: Suppose all processing times are a factor ~1.5 higher than in the Sc1... So, apply multiplicity cuts at L0 (NSPD<180 and NPU<80) to recover the same average for L1 time (0.5 ms), while preserving a reasonable signal efficiency. L0 and L1 thresholds readjusted to keep the nominal L0 and L1 retention rates. This is similar to assuming that we have only 2/3 of the CPUs available compared to baseline.
Use “realistic” data ( = full simulation, with noise, clusters per L1 board, etc.), including CPU processing time. (4.4M minimum bias events)
LHCb L1&DAQ review – April 29, 2003, CERN Massimiliano Ferro-Luzzi 9
L1 time and number of clusters
L1 time and number of clusters
Sum of all L1 TT+VELO clusters
L1 processing time (a.u.)
Sc1 (baseline)
Sc3 (multipl. cut), before re-scaling time
Considerable gain in average time and data size with small signal efficiency loss.
LHCb L1&DAQ review – April 29, 2003, CERN Massimiliano Ferro-Luzzi 10
Summary and OutlookSummary and Outlook
There seem to be no offline-selected signal events above a certain multiplicity
L0*L1 quite sensitive to multiplicity cut
Multiplicity cuts offer possibility to reduce average and tails ofprocessing time in L1&DAQ farm,
amount of data we put through the network
with little loss on signal efficiency (to be confirmed by multichannel study)
SPD + Pile-Up Veto (at L0) and Velo (at L1) multiplicities can be used as an additional handle to counter “unexpected adversity” from mother nature
Various scenarios under study
LHCb L1&DAQ review – April 29, 2003, CERN Massimiliano Ferro-Luzzi 11
Look at L1 Time for various multiplicity
cuts on VELO clusters
Look at L1 Time for various multiplicity
cuts on VELO clusters
Make as if applying a cut at the “entrance” of L1 processing:
if #L1Clusters>value then L1-no and set L1 time = 0.
Brunel v17r4 L0-yes MBIA
LHCb L1&DAQ review – April 29, 2003, CERN Massimiliano Ferro-Luzzi 12
L1L0 Global Performance vs multiplicity
cuts
L1L0 Global Performance vs multiplicity
cuts
Still Brunel v16r4 (Xmas)
Apply both PU and SPD multiplicity cuts
Rescale L0 thresholds and L1global to get proper MBIA retentions
Only 1 physics channel looked at…
…this is just a start.
LHCb L1&DAQ review – April 29, 2003, CERN Massimiliano Ferro-Luzzi 13
Same but with additional Velo
clusters cut
Same but with additional Velo
clusters cut
Apply a cut on number of Velo clusters at entrance of L1: < 2000.
LHCb L1&DAQ review – April 29, 2003, CERN Massimiliano Ferro-Luzzi 14
L1 versus VELO clustersL1 versus VELO clustersFrom Frederic Teubert @LHCC
BB00ss D Dss
--KK++
What if the Minimum What if the Minimum
Bias Bias multiplicity is multiplicity is wrongwrong??
What if the Minimum What if the Minimum
Bias Bias multiplicity is multiplicity is wrongwrong??
Subdivide the Subdivide the Minimum Minimum
Bias sample in:Bias sample in:
tracks < 40tracks < 40
40 < tracks < 70 40 < tracks < 70
70 < tracks70 < tracks
Subdivide the Subdivide the Minimum Minimum
Bias sample in:Bias sample in:
tracks < 40tracks < 40
40 < tracks < 70 40 < tracks < 70
70 < tracks70 < tracks
If the If the multiplicity multiplicity isis ~70% ~70%
higherhigher, the , the signal signal efficiency efficiency
isis reduced reduced by by ~30%.~30%.
If the If the multiplicity multiplicity isis ~70% ~70%
higherhigher, the , the signal signal efficiency efficiency
isis reduced reduced by by ~30%.~30%.
LHCb L1&DAQ review – April 29, 2003, CERN Massimiliano Ferro-Luzzi 15
L0>0L0>0
BBd d BBs s D Ds s
BBs s (()) BBs s (ee)(ee) BBdd K* K*
L0>0 && L1>0L0>0 && L1>0 L0>0 && L1>0 && SEL>0L0>0 && L1>0 && SEL>0
L1 versus VELO clustersL1 versus VELO clusters Thanks to Frederic Teubert
BBs s D Ds s KK
LHCb L1&DAQ review – April 29, 2003, CERN Massimiliano Ferro-Luzzi 16
L1 vs VELO clustersL1 vs VELO clusters Thanks to Frederic Teubert
Status at LHCCStatus at LHCC Velo < 2.2k clustersVelo < 2.2k clusters
LHCb L1&DAQ review – April 29, 2003, CERN Massimiliano Ferro-Luzzi 17
L1 Time versus Velo
Clusters multiplicity
L1 Time versus Velo
Clusters multiplicity
LHCb L1&DAQ review – April 29, 2003, CERN Massimiliano Ferro-Luzzi 18
Two scenarios: L1 time and number of clustersTwo scenarios: L1 time and number of clusters
Sum of all L1 TT+VELO clustersScaled L1 time (s)
baseline
Robustnesscheck
(Note: here real L1veloclusters; previous slides, offline Velo clusters)
What impact on farm design ?What impact on farm design ?