News on PXD Readout Issues Sören Lange, Universität Gießen, Belle II DAQ/TRG, KEK, 七夕, 平成 21

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) Outline qData Compression Algorithm reminder: input data rate ~27 GB/s we must achieve factor reduction by algorithm on FPGA qMost important piece of the algorithm: online track finder and track fitter qfinal algorithm results for Panda PIX+SVD (full digitization of simulated hits, optimized Hough space, „what is the best that we can achieve?“, for thesis of David Münchow, to be submitted Aug 20) qfirst algorithm results for Belle II PXD+SVD (Hough Space non-optimized yet) qNote: difference between Giessen and Munich approach q we assume xyz coordinates from clustering (no rphi or rz slices) q helix approach: we would like to get fitted momentum and not only straight line extrapolation

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) Online Track Finder and Track Fitter Step #1: Conformal Map for Track Finding inverse space (lever arm constraint by PXD hits i.e. highest resolution) real spaceconformal space

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) qIMPORTANT: our Hough space is filled with conformal coordinates! qfix point arithmetics (instead of floating point) q24 bit (in division and multiply 48 bit) qHough space 512×512 qlookup table for sinus(): 128 x 16 bit qp z reconstruction is a 2 nd Hough transform helix in xz space is a sinus function Online Track Finder and Track Fitter Step #2: Hough Transform for Track Fitting Parameter 1 Parameter 2 sinus form z x pTpT pzpz

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) Panda PXD+SVD geometry Full digitization switched on in simulation y / cm x / cm y / cm z / cm

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) Online Track Finder and Track Fitter Momentum Resolution p=0.3 GeVp=1 GeVp=3 GeV 2.9% (sigma) 8.9% (sigma) 2.9% (sigma) Final results for Panda (for PDX, SVD + Straw Tube Tracker, · 30 hits on track)  Efficiency ~10% lower than offline, over whole pT range  Momentum resolution factor ~2.5 worse than offline  Outlier hits ~6.2% (not assigned to any track)

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) Momentum Resolution p=100 MeV, MVD Stand-Alone (no STT) (also Test for Belle II Pixel Detector Trigger) where is the peak? p / GeV/c

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) Tracks with p=100 MeV, MVD only Real Space xyz Hits

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) Online Track Finder finds detector geometry

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) Belle II PXD+SVD Simulated Hits simulated hit data provided by Zdynek Drasal digitization not switched on yet y / mm x / mm z / mm r / mm

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) Results, Belle II PXD+SVD, Momentum Resolution, Stand Alone Tracking (0,0,0) required ¸ 3 hits required 2 issues: 1. Efficiency only ~25% because Hough space non-optimized (see next slide) Hough peak cut >4 entries will be reduced in the next step 2. Why is reconstructed momentum wrong? (see next slide) preliminary p=1 GeV/c Hough space non-optimized

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen)  there is a peak in Hough space in every event (if ¸ 3 hits) qoptimize efficiency we must find the peak (i.e. adjust threshold) qoptimize p T fitted there are tilted (diagonal) structures in Hough space (different from Panda case) > shifts peak center-of-gravity > gives wrong pT > checking ongoing (maybe from tilted SVD modules?) Hough Space event-by-event

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) How can we buffer for 5 seconds? qCompute node has only 2 GB DDR2 RAM per 1 FPGA  Per 1 FPGA, this is sufficient to buffer · 3 30kHz qHow can we buffer longer? > ATCA „memory module“? qe.g. RadiSys ATCA-4500 qProcessor Xeon L GHz quad-core Red Hat Enterprise Linux qup to 64GB DDR3 RAM (8 DIMM array) qOfficial price quotation from german RadiSys subsidiary 3.929,- US$ without RAM

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) Summary qA track finder and track fitter based on lookup tables and fixed point arithmetics was tested. qAnsatz: Hough space is filled with conformal coordinates qFor Panda the efficiency is ~10% worse than offline and the p T resolution is factor worse qFor low p T tracks, the algorithm finds the detector geometry (i.e. it connects hits in one layer) qThere are differences in Hough space between Panda and Belle II (probably of modules not parallel to z-axis or y-axis) to be checked

Backup Slides

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) Compute Node (CN) Concept q5 x VIRTEX4 FX-60 FPGAs qeach FPGA has 2 x 300 MHz PowerPC qLinux (open source version) qalgorithm programming in VHDL (XILINX ISE 10.1) qATCA (Advanced Telecommunications Computing Architecture) with full mesh backplane point-to-point connections on backplane from each CN to each other CN, i.e. no bus arbitration  8 optical links (connected to RocketIO at FPGA) · 6.5 Gbits/s q5 x Gigabit Ethernet qATCA management (IPMI) by add-on card

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) Compute Node (CN) Version 1.0

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) ATCA Shelf

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) Panda Silicon Vertex Detector q4 barrels & 8 disks qinner layers pixels qouter layers strips qforward mixed qSilicon Pixel Detectors q120 modules q100x100 μm2 pixel size q~10 7 readout channels qmaximum rate <10 MHits s–1 per module qRadiation length ~1% X0 per layer qSilicon Strip Detectors q400 modules q~0.5 m 2 active area q7x10 4 readout channels

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) Panda Silicon Vertex Detector q4 barrels & 8 disks qinner layers pixels qouter layers strips qforward mixed qSilicon Pixel Detectors q120 modules q100x100 μm2 pixel size q~10^7 readout channels qmaximum rate <10 MHits s–1 per module qRadiation length ~1% X0 per layer qSilicon Strip Detectors q400 modules q~0.5 m^2 active area q7x10^4 readout channels qFull digitization switched on in simulation

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) Panda Online Track Finder and Track Fitter Conformal Map + Hough Transform Momentum Resolution 7.3% 5.5% p / GeV/c Fit function p4+p3*exp(p1+p2*x) Preliminary not linear

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) MVD+STT – PandaRoot Momentum Resolution     –  5.5%7.3% online track finder Stefano Spataro, Panda Meeting March 2009

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) Panda Online Track Finder, Efficiency (# of found tracks) area in gaussian peak reconstructed momentum within § 10% of generated momentum efficiency decreases for higher momentum (same as offline) Preliminary p / GeV/c

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) MVD+STT – PandaRoot Tracking Efficiency     –  Online Track Finder  (p)= –0.0214·p Stefano Spataro, Panda Meeting March 2009

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) Quality of Online Track Finder Fake Tracks and Outlier Hits p=1 GeV/c average ' 6.2% Fake rate relative # of tracks (of all found tracks) with p reconstructed outside § 10% of p generated Number of Hits not assigned to a found track p / GeV/c

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) Conformal Map for Track Finding Particular case: reference point for all tracks on one track x 0, y 0 = this point this track missing hits!

PXD Issues, Belle II DAQ/TRG Session Sören Lange (Giessen) Step #1: Conformal Map for Track Finding Case: reference point for all tracks on one track x 0, y 0 = this point this track missing hits! If one hit on track is used for conformal transformation, also a track not pointing (0,0,0) (e.g. secondary from  decay) generates a straight line in conformal space If one hit on track is used for conformal transformation, also a track not pointing (0,0,0) (e.g. secondary from  decay) generates a straight line in conformal space