1 / 21

Upgrade Letter of Intent High Level Trigger Thorsten Kollegger

Upgrade Letter of Intent High Level Trigger Thorsten Kollegger. ALICE | Offline Week | 03.10.2012. Requirements. Focus of ALICE upgrade on physics probes requiring high statistics: sample 10 nb -1 Online System Requirements

latona
Télécharger la présentation

Upgrade Letter of Intent High Level Trigger Thorsten Kollegger

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. Upgrade Letter of IntentHigh Level TriggerThorsten Kollegger ALICE | Offline Week | 03.10.2012

  2. Requirements Focus of ALICE upgrade on physics probes requiring high statistics: sample 10 nb-1 Online System Requirements Sample full 50kHz Pb-Pb interaction rate (current limit at ~500Hz, factor 100 increase)  ~1.1 TByte/s detector readoutHowever: storage bandwidth limited to ~20 GByte/s many physics probes have low S/B: classical trigger/event filter approach not efficient ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

  3. Slide from Karel Safarik • Main physics topics, at the LHC uniquely accessible with the ALICE detector: • measurement of heavy-flavour transport parameters: • diffusion coefficient – azimuthal anisotropy and RAA • in-medium thermalization and hadronization – meson-baryon • mass dependence of energy loss – RAA • study of QGP properties via transport coefficients (h/s, q) • J/y , y’, and cc states down to zero pt in wide rapidity range • yields and transverse momentum spectra – RAA, elliptic flow • density dependence – central vs. forward production • statistical hadronization vs. dissociation/recombination ˆ Physics Motivation ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

  4. Physics Motivation Slide from Karel Safarik • measurement of low-mass and low-ptdi-leptons • chiral symmetry restoration – vector-meson spectral function • disappearance of vacuum condensate and generation of hadron masses • QGP thermal radiation – low-mass di-lepton continuum • space-time evolution of the QGP – radial and elliptic flow of emitted radiation • Jet quenching and fragmentation • jet energy recuperation at very low pt • heavy-flavourtagged jets, gluon vs. quark induced jets • heavy-flavour produced in fragmentation • particle identified fragmentation functions • Heavy-nuclear states • high statistics mass-4 and -5 (anti-)hypernuclei • search for H-dibaryon, Ln bound state, etc. ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

  5. Why not triggering? Slide from Luciano Musa Triggering on D0, Ds and Λc (pT>2 Gev/c)  ~ 36 kHz@50kHz rate... ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

  6. Strategy Data reduction by (partial) online reconstruction and compression Store only reconstruction results, discard raw data Demonstrated with TPC clustering since Pb-Pb 2011 Optimized data structures for lossless compression Algorithms designed to allow for offline reconstruction passes with improved calibrations  Implies much tighter coupling between online and offline reconstruction software ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

  7. Event Size Expected data sizes for minimum bias Pb-Pb collisions at full LHC energy ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

  8. TPC Data Reduction First steps up to clustering on FEE/FPNs (RORC FPGA)Further steps require full event reconstruction on EPNs, pattern recognition requires only coarse online calibration ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

  9. TPC Data Reduction Float to Fixed-Point convertion, size according to detector resolution Reduction of data size/cluster: 22 Byte -> 10 Byte ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

  10. TPC Data Reduction • Lossless data compression with Huffman code (entropy encoding) • Data members transformed to increase performance: • e.g. Padrow Number => Drow(i) = row(i) – row(i-1) • Entropy reduced from ~6 to 1.1 ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

  11. TPC Data Reduction Overall data size to tape reduced by factor 4.3 Used in Pb+Pb 2011, p+p 2012... standard ALICE data format Further reduction possible by transforming pad, time coordinates ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

  12. TPC Data Reduction First steps up to clustering on FEE/FPNs (RORC FPGA)Further steps require full event reconstruction on EPNs, pattern recognition requires only coarse online calibration ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

  13. Further data reduction • Discard clusters not assigned to tracks (or in the track vincinity) • Requires online calibration (at least coarse one) • Allows later offline re-production • Alternative: identify background clusters ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

  14. Processing Power Estimate for online systems based on current HLT processing power - ~2500 cores in ~200 nodes 108 FPGAs on H-RORCs for local preprocessing • TPC clusterfinding: 1 FPGA equivalent to ~80 CPU cores - 64 GPGPUs for tracking (NVIDIA GTX480 + GTX580) Scaling to 50 kHz rate to estimate requirements - ~ 250.000 cores additional processing power by FPGAs + GPGPUs 1250-1500 nodes in 2018 with multicores ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

  15. HLT TPC Tracking Algorithm implemented as multithreaded CPU and CUDA GPU version ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

  16. HLT TPC Tracking 3-fold speedup of GPU compared to optimized CPU version on 6 cores ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

  17. HLT Tracking Performance Active GPU Threads using Dynamic Scheduling time Consistency between GPU and CPU version of tracker threads Active GPU Threads: 67% ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

  18. HLT Tracking Performance Old HLT efficiency macro • Efficiency/Clone/Fake rate calculation • merged with PWG-PP/TPC code • Under review by TPC group ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

  19. Summary After the upgrade: Store only reconstruction results, discard raw data Requires online calibration Algorithms designed to allow for offline reconstruction passes with improved calibrations  Implies much tighter coupling between online and offline reconstruction software ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

  20. ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

  21. Backup - Processing Power Estimate of processing power based on scaling by Moore’s law However: no increase in single core clock speed, instead multi/many-core Reconstruction software needs to adapt to full use resources Picture from Herb Sutte: The Free Lunch Is Over A Fundamental Turn Toward Concurrency in Software Dr. Dobb's Journal, 30(3), March 2005 (updated) ALICE | Offline Week | 03.10.2012 | Thorsten Kollegger

More Related