310 likes | 598 Views
ALICE TPC Online Tracking on GPU. David Rohr for the ALICE Corporation 25.5.2010 Lisbon. Large Hadron Collider. Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary. ALICE TPC Online Tracking on GPU. The ALICE Experiment. Introduction Tracking Algorithm
E N D
ALICE TPC Online Tracking on GPU David Rohr forthe ALICE Corporation 25.5.2010 Lisbon
Large HadronCollider Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
The ALICE Experiment Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
Tracking Clusters Tracks Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
Real TPC pp-Event in Online Event Display Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
Simulated Heavy Ion Event Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
Tracks found in simulatedHeavy Ion Event Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
Tracks in simulated Central Heavy Ion Event Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
TPC Clusters divitedinto Slices Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
The ALICE TPC Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
TPC Trackerand Merger Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
One ALICE TPC Sector (Trakingisperformedrowbyrow) Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
Tracking Algorithm Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
Step I: NeighborsFinder (Fit beststraightlines) Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU dx+
Step II: Evolution (Keep coinciding links) Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
Step III: TrackletConstruction Green: SeedRed: Extrapolation Clusters closetotheextraplationpointaresearched Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
Neighbours Finder on Real Data Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
Evolution Step on Real Data Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
TrackletConstruction on Real Data Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
TrackletSelection on Real Data Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
NVIDIA CUDA GPU Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
Parallel TrackletConstruction Trackletsareindependentandcanbeprocessedsimultaneously Becauseofdatalocalitythetrackletsareprocessedfor a commonRow Single Instruction Decoder Either parallel fittingorextrapolation Red: Initial SeedGreen: Extrapolation Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU CurrentRow
Initial GPU Tracker Performance (PeripheralPb-PB) (8 threads on Nehalem CPU with 8 virtual / 4 physicalcores) Focus on TrackletConstructionwhenoptimizing Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
Active GPU Threads forthe First Implementation Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU Active GPU Threads: 19% Colors representTrackletConstructorSteps: Black: Idle Blue: Track Fit Green: Track Extrapolation x-axis: threads y-axis: time
Active GPU Threads using Dynamic Scheduling Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU Active GPU Threads: 62% Colors representTrackletConstructorSteps: Black: Idle Blue: Track Fit Green: Track Extrapolation x-axis: threadsy-axis: time
Pipelining Initialization / Output on CPU, Tracking on GPU and DMA Transfer canoverlap Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU Time
Final Speedup (Central Heavy Ion) (CPU performance was doubledas a sideeffect) Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
Final Speedup in Contrastto Event Size (PP Mode: Special variant optimizedforsmallscaleevents) Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU
Tracking Efficiency Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary ALICE TPC Online Tracking on GPU CPU GPU
Summary Introduction Tracking Algorithm NVIDIA CUDA Tracking on GPU Results Summary Twofoldperformanceincrease on CPU 2.5-fold performanceincreasecomparedto CPU Tracking efficiencymatchesthe CPU tracker‘sefficiency A commonsourcecodeensuresmaintainability CPU still availableduringthe GPU tracking ALICE TPC Online Tracking on GPU Perspectiveforthe Future Next NVIDIA GPU generationFermi mayresult in anotherboost Track Merger maybeadaptedtorun on the GPU