1 / 7

ALICE HLT tracking running on GPU

ALICE HLT tracking running on GPU. S. Gorbunov 1 and I. Kisel 1,2 ( for the ALICE Collaboration ) 1 Kirchhoff Institute for Physics, University of Heidelberg, Germany 2 Gesellschaft für Schwerionenforschung mbH , Darmstadt, Germany. ALICE/FAIR Workshop

scolumbus
Download Presentation

ALICE HLT tracking running on GPU

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. ALICE HLT tracking running on GPU S. Gorbunov1 and I. Kisel1,2 ( for the ALICE Collaboration ) 1Kirchhoff Institute for Physics, University of Heidelberg, Germany 2 Gesellschaft für Schwerionenforschung mbH, Darmstadt, Germany ALICE/FAIR Workshop GSI, February 3, 2009

  2. TPC reconstruction scheme TPC slice 35 TPC slice 0 The TPC Slice Tracker is the most complicated algorithm: • combinatorial search • fit mathematics • the reconstruction time is crucial Cluster finder Cluster finder clusters clusters Slice tracker Slice tracker … … … slice tracks slice tracks slice tracks TPC Global merger TPC tracks Sergey Gorbunov, KIP

  3. Tracking algorithm: the Cellular Automaton method Neighbours finder 2. Composing of tracklets • For each TPC cluster find two (up&down) neighbours which compose the best line • One-to-one linked neighbours are grouped to the track segments 3. Construction of the track candidates 4. Final selection of tracks • Competition between tracks, no shared clusters allowed • Fit of trajectories • Search for the missed parts Sergey Gorbunov, KIP

  4. Use of parallel hardware: GPU NVIDIA GeForce GTX 280: • 30x8 general propose processors; pure calculations can be ~100 times faster than CPU • very parallel: || execution of branches, || memory access • CUDA language - a little extension of the C++ • fast access to the small portion of data (16k) at the time; no memory cache • single precision floating point • ONLY parallel calculations Sergey Gorbunov, KIP

  5. Porting HLT tracking code to the GPU Pb-Pb event 156 ms 1640 ms in work The algorithm evolution: • Maximal parallelisation of the CPU tracker. ( AliRoot ) • Stand-alone CPU tracker w/o ROOT. ( AliRoot->stand-alone ) • Developing of the efficient GPU code. ( stand-alone CPU-> stand-alone GPU ) • Making hybrid code. ( stand-alone GPU -> stand-alone GPU+CPU ) • Porting code back to AliRoot. (stand-alone GPU+CPU -> AliRoot ) Result - HLT TPC tracker: • Official svn code, compiles and runs offline and in the HLT framework • Satisfies to the ALICE coding rules (almost) • Can use the GPU device (not from AliRoot) • SAME source code for GPU and CPU, same result. • For a moment >10.5 times faster on GPU Sergey Gorbunov, KIP

  6. Running the ALICE HLT tracker on the GPU cluster at Frankfurt • speed-up: 10.5x GPU CPU • same code • same result CPU GPU Sergey Gorbunov, KIP

  7. Summary and plans • Summary: • The ALICE HLT tracking algorithm has been parallelised to use the GPU hardware. • The new tracker is as fast as before on CPU, and shows 10x speed-up on GPU. • The algorithm and the code are universal for GPU and CPU. • Commit to svn, running in the HLT • Plans: • Further speed-up for the GPU • Integration of the GPU tracker to the HLT framework. Sergey Gorbunov, KIP

More Related