002704717 001__ 2704717
002704717 005__ 20240818013635.0
002704717 0248_ $$aoai:cds.cern.ch:2704717$$pcerncds:CERN$$pcerncds:CERN:FULLTEXT$$pcerncds:FULLTEXT
002704717 0247_ $$2DOI$$9bibmatch$$a10.1007/s41781-020-00039-7
002704717 037__ $$9arXiv$$aarXiv:1912.09161$$cphysics.ins-det
002704717 035__ $$9arXiv$$aoai:arXiv.org:1912.09161
002704717 035__ $$9Inspire$$a1771853
002704717 041__ $$aeng
002704717 100__ $$aAaij, [email protected]$$tGRID:grid.420012.5$$uNikhef, Amsterdam$$vNikhef National Institute for Subatomic Physics, Amsterdam, The Netherlands
002704717 245__ $$aAllen$$bA high level trigger on GPUs for LHCb
002704717 246__ $$9arXiv$$aAllen: A high level trigger on GPUs for LHCb
002704717 269__ $$c2019-12-18
002704717 260__ $$c2020-04-30
002704717 300__ $$a11 p
002704717 500__ $$9arXiv$$a12 pages, 12 figures, submitted to Computing and Software for Big
Science
002704717 520__ $$aWe describe a fully GPU-based implementation of the first level trigger for the upgrade of the LHCb detector, due to start data taking in 2021. We demonstrate that our implementation, named Allen, can process the 40 Tbit/s data rate of the upgraded LHCb detector and perform a wide variety of pattern recognition tasks. These include finding the trajectories of charged particles, finding proton-proton collision points, identifying particles as hadrons or muons, and finding the displaced decay vertices of long-lived particles. We further demonstrate that Allen can be implemented in around 500 scientific or consumer GPU cards, that it is not I/O bound, and can be operated at the full LHC collision rate of 30 MHz. Allen is the first complete high-throughput GPU trigger proposed for a HEP experiment.
002704717 520__ $$9Springer$$aWe describe a fully GPU-based implementation of the first level trigger for the upgrade of the LHCb detector, due to start data taking in 2021. We demonstrate that our implementation, named Allen, can process the 40 Tbit/s data rate of the upgraded LHCb detector and perform a wide variety of pattern recognition tasks. These include finding the trajectories of charged particles, finding proton–proton collision points, identifying particles as hadrons or muons, and finding the displaced decay vertices of long-lived particles. We further demonstrate that Allen can be implemented in around 500 scientific or consumer GPU cards, that it is not I/O bound, and can be operated at the full LHC collision rate of 30 MHz. Allen is the first complete high-throughput GPU trigger proposed for a HEP experiment.
002704717 520__ $$9arXiv$$aWe describe a fully GPU-based implementation of the first level trigger for the upgrade of the LHCb detector, due to start data taking in 2021. We demonstrate that our implementation, named Allen, can process the 40 Tbit/s data rate of the upgraded LHCb detector and perform a wide variety of pattern recognition tasks. These include finding the trajectories of charged particles, finding proton-proton collision points, identifying particles as hadrons or muons, and finding the displaced decay vertices of long-lived particles. We further demonstrate that Allen can be implemented in around 500 scientific or consumer GPU cards, that it is not I/O bound, and can be operated at the full LHC collision rate of 30 MHz. Allen is the first complete high-throughput GPU trigger proposed for a HEP experiment.
002704717 540__ $$3preprint$$aCC-BY-4.0$$uhttp://creativecommons.org/licenses/by/4.0/
002704717 540__ $$3publication$$aCC-BY-4.0$$uhttp://creativecommons.org/licenses/by/4.0/
002704717 542__ $$3preprint$$dCERN$$g2019
002704717 542__ $$3publication$$dThe Author(s)$$g2020
002704717 65017 $$2arXiv$$ahep-ex
002704717 65017 $$2SzGeCERN$$aParticle Physics - Experiment
002704717 65017 $$2arXiv$$aphysics.ins-det
002704717 65017 $$2SzGeCERN$$aDetectors and Experimental Techniques
002704717 690C_ $$aCERN
002704717 690C_ $$aARTICLE
002704717 690C_ $$aLHCb_DP
002704717 693__ $$aCERN LHC$$eLHCb
002704717 700__ $$aAlbrecht, Johannes$$iINSPIRE-00259834$$tGRID:grid.5675.1$$uTech. U., Dortmund (main)$$vFakultät Physik, Technische Universität Dortmund, Dortmund, Germany
002704717 700__ $$aBilloir, Pierre$$iINSPIRE-00067095$$tGRID:grid.463935.e$$uParis U., VI-VII$$vLPNHE, Sorbonne Université, Paris Diderot Sorbonne Paris Cité, CNRS/IN2P3, Paris, France
002704717 700__ $$aBoettcher, Thomas$$iINSPIRE-00582163$$tGRID:grid.116068.8$$uMIT$$vMassachusetts Institute of Technology, Cambridge, USA
002704717 700__ $$aBrea Rodriguez, Alexandre$$tGRID:grid.11794.3a$$uU. Santiago de Compostela (main)$$vInstituto Galego de Física de Altas Enerxías (IGFAE), Universidade de Santiago de Compostela, Santiago de Compostela, Spain
002704717 700__ $$aVom Bruch, Dorothea$$jORCID:[email protected]$$tGRID:grid.463935.e$$uParis U., VI-VII$$vLPNHE, Sorbonne Université, Paris Diderot Sorbonne Paris Cité, CNRS/IN2P3, Paris, France
002704717 700__ $$aCampora Perez, Daniel [email protected]$$tGRID:grid.420012.5$$tGRID:grid.5012.6$$uNIKHEF, Amsterdam$$vNikhef National Institute for Subatomic Physics, Amsterdam, The Netherlands$$vFaculty of Science and Engineering, Maastricht University, Maastricht, The Netherlands
002704717 700__ $$aCasais Vidal, Adrian$$tGRID:grid.11794.3a$$uU. Santiago de Compostela (main)$$vInstituto Galego de Física de Altas Enerxías (IGFAE), Universidade de Santiago de Compostela, Santiago de Compostela, Spain
002704717 700__ $$aCraik, Daniel Charles$$iINSPIRE-00357145$$tGRID:grid.116068.8$$uMIT$$vMassachusetts Institute of Technology, Cambridge, USA
002704717 700__ $$aFernandez Declara, Placido$$tGRID:grid.9132.9$$tGRID:grid.7840.b$$uCERN$$vEuropean Organization for Nuclear Research (CERN), Geneva, Switzerland$$vDepartment of Computer Science and Engineering, University Carlos III of Madrid, Madrid, Spain
002704717 700__ $$aGligorov, Vladimir$$iINSPIRE-00260081$$tGRID:grid.463935.e$$uParis U., VI-VII$$vLPNHE, Sorbonne Université, Paris Diderot Sorbonne Paris Cité, CNRS/IN2P3, Paris, France
002704717 700__ $$aJashal, Brij Kishor$$jORCID:0000-0002-0025-4663$$tGRID:grid.4711.3$$uValencia U., IFIC$$vInstituto de Física Corpuscular, Centro Mixto Universidad de Valencia, CSIC, Valencia, Spain
002704717 700__ $$aKazeev, Nikita$$tGRID:grid.410682.9$$uHigher Sch. of Economics, Moscow$$vNational Research University Higher School of Economics, Moscow, Russia$$vYandex School of Data Analysis, Moscow, Russia
002704717 700__ $$aMartinez Santos, Diego$$iINSPIRE-00259733$$tGRID:grid.11794.3a$$uU. Santiago de Compostela (main)$$vInstituto Galego de Física de Altas Enerxías (IGFAE), Universidade de Santiago de Compostela, Santiago de Compostela, Spain
002704717 700__ $$aPisani, Flavio$$tGRID:grid.9132.9$$tGRID:grid.470193.8$$tGRID:grid.6292.f$$uCERN$$vEuropean Organization for Nuclear Research (CERN), Geneva, Switzerland$$vINFN Sezione di Bologna, Bologna, Italy$$vUniversità di Bologna, Bologna, Italy
002704717 700__ $$aQuagliani, Renato$$iINSPIRE-00453676$$tGRID:grid.463935.e$$uParis U., VI-VII$$vLPNHE, Sorbonne Université, Paris Diderot Sorbonne Paris Cité, CNRS/IN2P3, Paris, France
002704717 700__ $$aRangel, Murilo$$iINSPIRE-00024740$$tGRID:grid.8536.8$$uRio de Janeiro Federal U.$$vInstituto de Física, Universidade Federal do Rio de Janeiro (UFRJ), Rio de Janeiro, Brazil
002704717 700__ $$aReiss, Florian$$tGRID:grid.463935.e$$uParis U., VI-VII$$vLPNHE, Sorbonne Université, Paris Diderot Sorbonne Paris Cité, CNRS/IN2P3, Paris, France
002704717 700__ $$aSanchez Mayordomo, Carlos$$iINSPIRE-00454015$$tGRID:grid.4711.3$$uValencia U., IFIC$$vInstituto de Física Corpuscular, Centro Mixto Universidad de Valencia, CSIC, Valencia, Spain
002704717 700__ $$aSchwemmer, Rainer$$iINSPIRE-00260360$$tGRID:grid.9132.9$$uCERN$$vEuropean Organization for Nuclear Research (CERN), Geneva, Switzerland
002704717 700__ $$aSokoloff, Michael$$iINSPIRE-00127570$$tGRID:grid.24827.3b$$uCincinnati U.$$vUniversity of Cincinnati, Cincinnati, OH, USA
002704717 700__ $$aStevens, Holger$$iINSPIRE-00582780$$tGRID:grid.5675.1$$uTech. U., Dortmund (main)$$vFakultät Physik, Technische Universität Dortmund, Dortmund, Germany
002704717 700__ $$aUstyuzhanin, Andrey$$iINSPIRE-00392208$$tGRID:grid.410682.9$$tGRID:grid.35043.31$$uYandex Sch. Data Anal., Moscow$$uHigher Sch. of Economics, Moscow$$vNational Research University Higher School of Economics, Moscow, Russia$$vYandex School of Data Analysis, Moscow, Russia$$vNational University of Science and Technology MISIS, Moscow, Russia
002704717 700__ $$aVilasis-Cardona, Xavier$$iINSPIRE-00259644$$uU. Barcelona (main)$$uRamon Llull U., Barcelona
002704717 700__ $$aWilliams, Mike$$iINSPIRE-00342010$$tGRID:grid.116068.8$$uMIT$$vMassachusetts Institute of Technology, Cambridge, USA
002704717 773__ $$c7$$n1$$pComput. Softw. Big Sci.$$v4$$y2020
002704717 8564_ $$81777272$$s1175529$$uhttps://cds.cern.ch/record/2704717/files/s41781-020-00039-7.pdf$$yFulltext from publisher
002704717 8564_ $$81777453$$s1175529$$uhttps://cds.cern.ch/record/2704717/files/10.1007_s41781-020-00039-7.pdf$$yFulltext from Publisher
002704717 8564_ $$82217368$$s1610074$$uhttps://cds.cern.ch/record/2704717/files/1912.09161.pdf
002704717 8564_ $$82217369$$s10284$$uhttps://cds.cern.ch/record/2704717/files/Fig6b.png$$y00004 :
002704717 8564_ $$82217370$$s48736$$uhttps://cds.cern.ch/record/2704717/files/Fig10.png$$y00012 Efficiency of the 1-Track and 2-Track trigger lines when calculating the IP (see text for definition) from tracks fitted with the simple and parameterized Kalman filter, using the $\Bs\to\phi\phi$ sample. Varying the selection criteria of the IP $\chi^2$ results in rate and efficiency changes. The efficiency is calculated from subsets of the sample, the central value and error band correspond to the mean and standard deviation respectively.
002704717 8564_ $$82217371$$s7728$$uhttps://cds.cern.ch/record/2704717/files/Fig11.png$$y00013 Allen throughput on various GPUs with respect to their reported peak 32-bit FLOPS performance. The mean and standard deviation of 10 measurements with different sets of 1000 events each are shown in the figure, with the measurement setup as described in the text.
002704717 8564_ $$82217372$$s9080$$uhttps://cds.cern.ch/record/2704717/files/Fig12.png$$y00014 Throughput of the Allen sequence as a function of the SciFi raw data volume, which is proportional to the SciFi occupancy. The measurement setup is described in the text. For every data point, 1000 different events within the range of the SciFi raw data volume bin are processed. The GEC removing the 10\% busiest events was deactivated for these measurements.
002704717 8564_ $$82217373$$s11435$$uhttps://cds.cern.ch/record/2704717/files/Fig7.png$$y00009 PV reconstruction efficiency versus track multiplicity of the MC PV for minimum bias events. The track multiplicity distribution is overlaid as a histogram.
002704717 8564_ $$82217374$$s1138813$$uhttps://cds.cern.ch/record/2704717/files/Fig4.png$$y00003 Upgraded LHCb detector. The y-component of the magnetic field $B_y$ is overlaid to visualize in which parts of the detector trajectories are bent. The maximum $B_y$ value is 1.05~T.
002704717 8564_ $$82217375$$s13862$$uhttps://cds.cern.ch/record/2704717/files/Fig2.png$$y00001 Threads are grouped into blocks, forming a grid that executes one kernel on the GPU.
002704717 8564_ $$82217376$$s12838$$uhttps://cds.cern.ch/record/2704717/files/Fig3.png$$y00002 In the GPU-enhanced proposal for the upgraded LHCb data acquisition system x86 event building units receive data from the subdetectors and build events by sending and receiving event fragments over a 100G Infiniband (IB) network. The same x86 servers also host GPUs which process HLT1. Built events are sent to x86 event filter servers at a rate reduced by a factor 30 - 60 to process HLT2.
002704717 8564_ $$82217377$$s11940$$uhttps://cds.cern.ch/record/2704717/files/Fig1.png$$y00000 In the baseline proposal for the upgraded LHCb data acquisition system, x86 event building units receive data from the subdetectors and build events by sending and receiving event fragments over a 100G Infiniband (IB) network. The full data stream of built events is sent to x86 event filter servers to process both stages of the high level trigger.
002704717 8564_ $$82217378$$s10265$$uhttps://cds.cern.ch/record/2704717/files/Fig6c.png$$y00005 :
002704717 8564_ $$82217379$$s11189$$uhttps://cds.cern.ch/record/2704717/files/Fig6f.png$$y00008 : Caption not extracted
002704717 8564_ $$82217380$$s10396$$uhttps://cds.cern.ch/record/2704717/files/Fig6e.png$$y00007 : Caption not extracted
002704717 8564_ $$82217381$$s10997$$uhttps://cds.cern.ch/record/2704717/files/Fig6d.png$$y00006 :
002704717 8564_ $$82217382$$s9454$$uhttps://cds.cern.ch/record/2704717/files/Fig8.png$$y00010 Relative momentum resolution of tracks passing through the Velo, UT and SciFi detectors versus momentum for all signal samples combined. Points represent the mean, error bars the width of a Gaussian distribution fitted to the resolution in every momentum slice. The momentum distribution is overlaid as a histogram.
002704717 8564_ $$82217383$$s12197$$uhttps://cds.cern.ch/record/2704717/files/Fig9.png$$y00011 Muon identification efficiency versus momentum for tracks passing through the Velo, UT and SciFi detectors with respect to all reconstructible muons (explained in the text), for all signal samples combined. The momentum distribution is overlaid as a histogram.
002704717 960__ $$a13
002704717 981__ $$a2705531
002704717 980__ $$aARTICLE