Presentation is loading. Please wait.

Presentation is loading. Please wait.

WP8 : High Level Trigger John Baines. Tasks & Deliverables WP8 Tasks: 1.Optimize HLT Tracking software for Phase-I 2.Optimize Trigger Selections for Phase-I.

Similar presentations


Presentation on theme: "WP8 : High Level Trigger John Baines. Tasks & Deliverables WP8 Tasks: 1.Optimize HLT Tracking software for Phase-I 2.Optimize Trigger Selections for Phase-I."— Presentation transcript:

1 WP8 : High Level Trigger John Baines

2 Tasks & Deliverables WP8 Tasks: 1.Optimize HLT Tracking software for Phase-I 2.Optimize Trigger Selections for Phase-I 3.Upgrade Trigger Steering Software 4.Simulation studies to optimize selection strategy for Phase-II DeliverableDate Steering software upgradedDec 12 HLT Tracking optimizedMar 13 Selection software optimizedMar 13

3 Overview Primary focus : Phase 0 & Phase I : Upgrade tracking: – pat. rec. in high occupancy (> 25 p.u. events per b.c.) – Speed up the code to limit rise of execution time with occupancy – Possible use of GPU – target most time-consuming code – Integration of IBL, use of FTK information (hardware-based tracking running before HLT) Upgrade Steering Software: – Changes to track evolving computing hardware & sw incl. changes in offline & TDAQ sw – Provide needed additional flexibility needed : Full event/RoI processing, Handling of “pathological events” Combined L2 & EF, Provision for FTK & IBL Upgrade selections for L > 10 34 cm -2 s -1 – Tighter cuts, greater use of topological info – HLT selection chains for MC productions to optimize with L1 & detector upgrades For bulk of work, timescale same (or advanced) c.f. proposal – FTK slice tests already in 2012, need to be prepared for possible FTK deployment for 2014 – Prepare for IBL installed in 2013/14 shutdown – Prepare for > design occupancies in 2015

4 HLT Tracking First 13 months: Benchmarking existing ID & Muon tracking code Design of upgraded L2 ID software package Assessment of speedup from GPU

5 ID Tracking at LVL2 Trigger Cross-section view Level 1 Trigger RoI Data preparation data requests ATLAS detector raw data spacepoints LVL2 track finding by Hough transform: IdScan combinatorial track finding: SiTrack, offline track candidates Tracks Track fitting interaction vertex finding: ATLAS z-axis pp Hough transform in space

6 WP8 : Performance Benchmarking No pileup 10 34 pileup 2x10 34 pileup Have established test-bed to benchmark performance on MC datasets up to 2x10 34 cm -2 s -1 Measured performance: Efficiency of track reconstruction as a function of luminosity Execution time as a function of occupancy Identified areas for optimisation & improvement Next steps: Look at different RoI types : electron, jet, tau etc. Re-run with tracking optimisations Execution time v. no. spacepoints Top events 10 34 pileup Level-2 Inner Detector Tracking (Muon RoI) Level-2 Inner Detector Tracking (Muon RoI) top events

7 Optimisation of L2 ID tracking code Start with IDScan Zfinder Optimize Effic. & Timing Optimize Code Effic. v. value for various Zfinder parametersTiming v. value for various Zfinder parameters

8 Upgraded L2 ID Software Package TrigSteering The unified algorithm Data Provider Track Fitting Monitoring TRT Track ext. > ITrigL2PattRecoStrategy m_findTracks(RoI, data) : HLT::ErrorCode m_findTracks(data) : HLT::ErrorCode Strategy AStrategy BStrategy C > ITrigL2CommonZFinder m_findZvertices(RoI, data) : HLT::ErrorCode m_findZvertices(data) : HLT::ErrorCode HLT_Algo AlgTool Configurable Pat. Rec. strategies > ICombTrackFinder m_findTracks(seeds, tracks&) AlgTool Schematic Design of L2STAR Algorithm New L2 ID Package: Framework providing : Interface to Steering Data Preparation Track Fitting & Propogation Monitoring Configurable Pattern Recognition Allows : upgraded sw to be developed independantly Provides single algorithm framework for L2 ID Flexibility to develop & adapt strategies for: High occupancy environment Inclusion of IBL Strategy for tracking using FTK i/p

9 ~factor 10 speed up for Fermi GPU GPUCPU Use of GPU in the HLT Measure possible speed-up of L2 code on GPU c.f. CPU Ported HLT code to GPU: – Zfinder – Track Fitter – Data preparation Measure execution times c.f. CPU Factor 35 speed-up for L2 Zfinder running on GPU GPU time almost flat Also ported Data Preparation Next steps: Increase parallelisation of fitter Port Pixel clustering to GPU  Complete tracking chain on GPU 1000 2000 3000 Tracks per event 80 70 60 50 40 30 20 10 Time (ms) With WP9

10 Test of client-server architecture Implemented test based on combinatorial seeding code GPU: Tesla C2050 (Fermi chip), CPU: 2.4 GHz Westmere Data transfer: Host-to-GPU ~ 4.0 Gb/s (limit 6Gb/s) Time of data transfer to GPU: ~ 0.34 ms for ~ 2000 spacepoints Factor 29 speedup Separation of CUDA GPU code from athena C++ code using client-server technology Provides a set of high-level routines in athena : findTracks, findVertices Athena Service ComputeSvc sends the task to the ComputeServer which, in turn, starts the corresponding kernel on GPU

11 Muon Tracking One challenge is to reduce execution time for Pat. Rec. due to large no. combinations in high occupancy events Have identified sources of timeouts in EF Substantial speed-up acheived Execution time per RoI Original code Optimized code

12 Next 12 Months Extend benchmarking to cover all trigger chains Continue optimization of components of L2 ID tracking code – incl. Specific optimizations for specific chains Complete Implementation of Tracking strategies in New L2 Package. Continue to benchmark & optimize EF ID and Muon tracking code Implement complete tracking chain on GPU Define Trigger Selections for MC productions Start upgrade of Steering software – add flexibility needed for upgrade studies

13 Issues Late confirmation of funding => delay to filling NP2 Original start date 15 th Nov 2010 (post descope). Will now start 1st June 2011  6.5 months delay to start of steering upgrade work  Consequent delay to work relying on these upgrades

14 Milestones Milestone DescriptionOrigin al Date Target Date Act ual Date Status M8.1First GPU measurements : Zfinder & fitter NewDec 2010 Complete M8.2L2 Zfinder optimizedNewSep 11In progress M8.3HLT Tracking Code updated for IBL Mar 11Dec 11Delayed – lack of availiability of offline code (external) M8.4Trigger Selections Defined for MC Productions Dec 11 M8.5Steering Code upgradedMar 12Dec 12Delayed – late start of New Posts M8.6GPU speed-up measured for complete L2 tracking chain Dec 12Mar 12 M8.7HLT Selections updated for upgraded LVL1 Dec 12Mar 13Delayed – late start of New Posts M8.8HLT Tracking code optimised for Phase-I Mar 13

15 Summary Very good progress over first 12 months: – Benchmarking test-bed in place - used to target code of optimization – Improvements made to ID tracking code IDScan Zfinder EF Muon Pat. Rec – Encouraging results from GPU studies: Factor 35 speedup for Zfinder Factor 10 speedup for track fitter Late approval of funds for new posts => delays in start of core s/w work – Post will start 1 st June


Download ppt "WP8 : High Level Trigger John Baines. Tasks & Deliverables WP8 Tasks: 1.Optimize HLT Tracking software for Phase-I 2.Optimize Trigger Selections for Phase-I."

Similar presentations


Ads by Google