Distribution of the difference between the number of clusters identified in each event by the GPU-accelerated implementation of the Topo-Automaton Clustering algorithm under the approximation that the calorimeter noise is constant for all events and by the standard CPU implementation of calorimeter Topological Clustering, for two samples corresponding to Monte Carlo simulated ![]() ![]() ![]() ![]() |
![]() png eps |
Distribution of the number of clusters per event that are identified by the GPU-accelerated implementation of the Topo-Automaton Clustering algorithm under the approximation that the calorimeter noise is constant for all events (solid and dotted lines) or by the standard CPU implementation of calorimeter Topological Clustering (dashed and dash-dotted lines) and that have not been successfully matched to a cluster in the other implementation. Cluster matching is done through a variant of the Gale-Shapley algorithm![]() ![]() ![]() ![]() |
![]() png eps |
Distribution of the number of cells per cluster that have been assigned to clusters that do not match between the GPU-accelerated implementation of the Topo-Automaton Clustering algorithm under the approximation that the calorimeter noise is constant for all events (solid and dotted lines) and by the standard CPU implementation of calorimeter Topological Clustering (dashed and dash-dotted lines). Cells in a cluster that do not belong to a cluster in the other implementaton are included in this number. Cluster matching is done through a variant of the Gale-Shapley algorithm![]() ![]() ![]() ![]() |
![]() png eps |
Distribution of the differences between the pseudo-rapidities ( ![]() ![]() ![]() ![]() ![]() |
![]() png eps |
Distribution of the speed-up made possible by GPU acceleration in each event, which corresponds to the ratio between the time taken to process it by a single-threaded execution of the standard CPU implementation of calorimeter Topological Clustering to that taken by the GPU-accelerated implementation of the Topo-Automaton Clustering algorithm. The samples that were used correspond to Monte Carlo simulated ![]() ![]() ![]() ![]() ![]() ![]() |
![]() png eps |
Fraction of events in which the L1 EM trigger found within the data selected by the minimum bias scintillators (MBTS) as a function of the date of data taking with stable beam collisions is shown. Stable L1 operation is achieved within 10%. This plots shows all the good collision runs taken between March 31 and April 18, 2010. A clear change is seen around April 10 arising from the updates to the L1 timing constants. |
![]() gif eps |
L1 EM trigger rate as a function of the threshold for cosmic and collision candidates. The rates are averaged over the run period and normalised to correspond to one pair of colliding bunches are shown for all events passing the EM3 trigger for collision events from the 900 GeV and 7 TeV data taking periods. In addition, the rate for one run taken during cosmic ray data taking is given. |
![]() gif eps |
L1 efficiency for the trigger selecting EM clusters above 2, 3 or 5 counts (1 count is ~1 GeV) as a function of the raw (uncalibrated) offline cluster ET. The turn-on is shown for data (markers) and MC simulation (lines). The turn-on in data is happening at slightly lower ET values compared to MC and the shape is well modeled. The L1 selected events at low ET values arise from offline clusters with nearby energy deposits. This is due to the larger region in space of L1 clusters compared to offline ones. |
![]() gif eps |
L1 efficiency for the trigger selecting EM clusters above 5 counts (1 count is ~1 GeV) as a function of the raw (uncalibrated) offline cluster ET. The turn-on is shown for data (markers) and MC simulation (lines). The turn-on in data is happening at slightly lower ET values compared to MC and the shape is well modelled. The L1 selected events at low ET values arise from offline clusters with nearby energy deposits. This is due to the larger region in space of L1 clusters compared to offline ones. |
![]() gif eps |
L1 efficiency for the trigger selecting EM clusters above 3 counts ( ~3 GeV) as a function of the raw (uncalibrated) offline cluster ET. This plot shows nicely she sharpening of the turn-on after the L1Calo timing correction. It also shows that the plateau value now reaches 100%. |
![]() gif eps |
L1 Rates normalised to L=1027cm-2s-1. The errors shown are statistical only. The error on the luminosity is at this point in time ~20%. For example EM2 means EM clusters found above 2 counts or at least 3 counts (1 count ~1 GeV) |
![]() gif |
The ET resolution for online (HLT) topo-clusters with respect to offline topo-clusters with ET > 3 GeV. The online and offline clusters are both hadronically calibrated. Both the 2015 data and 2016 data are comprised of bunch trains containing 72 filled bunches. No explicit BCID-based event selection has been applied and so all bunch crossings (BCs) are represented. The improvement in the ET resolution for online topo-clusters with respect to offline topo-clusters in 2016 is due to the introduction cell-level, BCID/<> based energy corrections. These corrections account for pedestal changes that arise due to out-of-time pile-up and particularly affect the first bunch crossings in each bunch train. Similar corrections were already being applied offline. | ![]() pdf eps gif |
The ET resolution for online (HLT) topo-clusters with respect to offline topo-clusters with ET > 3 GeV. The online and offline clusters are both hadronically calibrated. Both the 2015 data and 2016 data are comprised of bunch trains containing 72 filled bunches. Only the events corresponding to the first 12 bunch crossings (BCs) in each bunch train are considered here. The improvement in the ET resolution for online topoclusters with respect to offline topo-clusters in 2016 is due to the introduction cell-level, BCID/<> based energy corrections. These corrections account for pedestal changes that arise due to out-of-time pile-up and particularly affect the first bunch crossings in each bunch train. Similar corrections were already being applied offline. | ![]() pdf eps gif |
The ET resolution for online (HLT) topo-clusters with respect to offline topo-clusters with ET > 3 GeV. The online and offline clusters are both hadronically calibrated. Both the 2015 data and 2016 data are comprised of bunch trains containing 72 filled bunches. The solid lines show the ET resolution for all events, whereas the dotted lines show the ET resolution for the sub-set of events that correspond to the first 12 bunch crossings in each bunch train. The improvement in the ET resolution for online topo-clusters with respect to offline topo-clusters in 2016 is due to the introduction cell-level, BCID/<> based energy corrections. These corrections account for pedestal changes that arise due to out-of-time pile-up and particularly affect the first bunch crossings in each bunch train. Similar corrections were already being applied offline. | ![]() pdf eps gif |
L2 e/gamma clusters transverse energy distribution ATLAS trigger electromagnetic clusters transverse energy distribution for 7 TeV collisions (668ub-1). Online monitoring plot showing the effect of cosmic ray events in addition to collision events. By selecting collisions only events, based on beam presence, the overall shape has a fair agreement with simulated data. |
![]() eps |
EF e/gamma clusters eta distribution ATLAS trigger electromagnetic clusters eta coordinate for 7 TeV collisions (668ub-1). Beam presence is used to select real collision events. Important discrepancy verified in the services gap region (eta=1.5) are expected due to lack of information input to the hardware based first trigger level. This information will be introduced to the hardware trigger after more detailed collision data based calibrations can be performed. |
![]() eps |
ET spectrum for data collected in 2009 with 900 GeV collisions (~9 mub-1 of stable beam data) and data collected up to April 17 in 2010 with 7 TeV collisions (~400 mub-1 of stable beam data) is shown together with their Monte-Carlo predictions. |
![]() gif eps |
Comparison of the distributions of the shower shape Reta calculated at EF for data and simulation for EF clusters matched to an offline electron candidate. Reta calculated by the ratio of the energy deposit in 3x7 cells (corresponding to 0.075 x 0.175 in eta x phi) over 7x7 cells in the second EM layer. Note, the value of Reta can have values above one as cell energies can be negative due to the electronic shaping function used in LAr. The shift in the distribution to slightly lower values in data mainly arises from cross-talk. |
![]() gif eps |
Difference in Reta between values found by the HLT and offline reconstruction for L2/EF clusters matched to offline electron candidates. Reta calculated by the ratio of the energy deposit in 3x7 cells (corresponding to 0.075 x 0.175 in eta x phi) over 7x7 cells in the second EM layer. The broader distribution found at L2 arises from small differences in the clustering algorithms which results in slightly different cluster positions. |
![]() gif eps |
Distribution of shower shape in 1st EM layer Eratio=(E1max-E2max)/ (E1max+E2max) at EF for data and MC for EF clusters matched to offline electron candidates. The data shows a shift in the distribution to slightly lower values which hints at missing material in the MC. |
![]() gif eps |
Difference between the variable Eratio=(E1max-E2max)/ (E1max+E2max) calculated in the 1st EM layer found by L2 and EF w.r.t. the offline for L2 and EF clusters matched to offline electron candidates. |
![]() gif eps |
Distribution of the hadronic leakage ET(had) / ET(EM) at L2 for data and MC for L2 clusters matched to offline electron candidates. A good agreement between data and MC observed with a slightly sharper distribution around zero for observed in data. |
![]() gif eps |
Difference in ET between the values found by the EF and the offline reconstruction for EF clusters matched to offline electron candidates. Note the trigger uses the DSP cell energies as input whereas the offline recalculates optimal filtering coefficients in the offline. The re-calculation in the offline is only a start-up feature. |
![]() gif eps |
Event display shows the L2 quantities in the X-Y and R-z plan. The 3d plot shows the L1 tower information. In this event the L1 ET passes the 32 GeV threshold. L2 cluster: ET = 34.4 GeV, eta = -0.42, phi = 0.31; L2 track: pT = 24.9 GeV |
![]() png |
Comparison of the eta difference between the track extrapolated into the first EM layer and the barycentre of the cell energies in this layer calculated at L2 for data and simulation for trigger objects matched to offline electron candidates. The contribution of electrons from conversions found in the MC is shown separately. A reasonable agreement between data and MC is found |
![]() gif eps |
Comparison of the eta difference between the track extrapolated into the first EM layer and the barycentre of the cell energies in this layer calculated at EF for data and simulation for trigger objects matched to offline electron candidates. The contribution of electrons from conversions found in the MC is shown separately. A reasonable agreement between data and MC is found |
![]() gif eps |
Difference between the cluster and track eta positions found by the HLT and the offline. The distributions are shown for L2 and EF separately for L2 and EF objects matched to offline electron candidates. A very good agreement is found between EF and offline quantities,. The agreement is less good between L2 and offline mainly due to L2 tracking algorithm, which is less sophisticated than the offline one due to the timing constraints in the online. |
![]() gif eps |
Processing time per event for the L2 tracking algorithm. These measurements are taken in the online farm. |
![]() gif eps |
Number of calorimeter clusters reconstructed using the standard CPU cell clustering algorithms and the same logical algorithm ported to GPU. The blue line represents the CPU standard CaloTopoClusterMaker algorithm. The red dashed line represents the GPU cell clustering based on a Cellular Automaton algorithm. In both cases the clusters are then sent to the CPU cluster splitting and follow the same reconstruction sequence. The results presented were obtained after the complete Trigger Clustering execution. The data sample used consist of QCD di-jet events with leading-jet transverse momentum above 20 GeV and a fixed number of 40 simultaneous interactions per bunch-crossing. |
![]() png pdf |
Transverse energy of the calorimeter clusters reconstructed using the standard CPU cell clustering algorithms and the same logical algorithm ported to GPU. The blue line represents the CPU standard CaloTopoClusterMaker algorithm. The red dashed line represents the GPU cell clustering based on a Cellular Automaton algorithm. In both cases the clusters are then sent to the CPU cluster splitting and follow the same reconstruction sequence. The results presented were obtained after the complete Trigger Clustering execution. The data sample used consist of QCD di-jet events with leading-jet transverse momentum above 20 GeV and a fixed number of 40 simultaneous interactions per bunch-crossing. |
![]() png pdf |
Number of jets reconstructed from clusters created using the standard CPU cell clustering algorithms and the same logical algorithm ported to GPU. The blue line represents the CPU standard CaloTopoClusterMaker algorithm. The red dashed line represents the GPU cell clustering based on a Cellular Automaton algorithm. In both cases the clusters are then sent to the CPU cluster splitting and follow the same reconstruction sequence. The results presented were obtained after the complete jet chain execution. The data sample used consist of QCD di-jet events with leading-jet transverse momentum above 20 GeV and a fixed number of 40 simultaneous interactions per bunch-crossing. |
![]() png pdf |