Tracking and compression techniques for ALICE HLT Anders Strand Vestbø
The ALICE experiment at LHC
The ALICE High Level Trigger (HLT) Estimated data rate (Central Pb-Pb, TPC only) 200 Hz * 75 MB = ~15 GB/s (Event rate)*(event size) Why? How? Detectors Constrains on data rate to mass storage = 1.2 GB/s DAQ HLT Mass storage Available data rate needs to be scaled down by an order of magnitude! Event selection based on software trigger Efficient data compression
Tasks Online TPC pattern recognition Online TPC data compression All data has been simulated with the AliROOT framework Event generator: HIJING parameterization, dn ch /d TPC microscopic simulator: Full detector response chain
Online TPC pattern recognition Multiplicities The algorithms Performance Computing requirements
Multiplicities and occupancies What to expect? Extreme assumption: dn ch /dy=8000 Extrapolations from RHIC: dn ch /dy=2000-4000 TPC occupancy for different multiplicities
TPC tracking methods Cluster finding Reconstruct space points from 2D clusters Track seed finder Track candidates Cluster finder / fitter Cluster shape is a function of track parameters Sequential tracking Iterative tracking Space points Track finder fitter on tracks Track reconstruction Connect space points into tracks and fit them to a model (helix)
Glossary OFFLINE: Results obtained with the official reconstruction chain in ALICE (used for the offline analysis) - Based on the Kalman filtering approach HLT (High Level Trigger): Results obtained with the algorithms/framework developed here
Sequential tracking the algorithms Input: ADC-sequences above threshold Cluster Finder - Simple sequence matching between neighboring pads - 2 lists in memory principle: current and previous pad(s) - Centroids calculated as weighted mean of ADC-values Simple deconvolution scheme: Split clusters at local minima Track Finder Input: 3D space points - Conformal mapping Real space Conformal space x' = x - x v r 2 y' = y - y v r 2 r 2 = (x-x v ) 2 + (y-y v ) 2 - Follow-your-nose Build tracks from outer to inner TPC-radius
Tracking efficiency Tracking efficiency Pt / pt [%] Pt / pt [%] [%] Sequential tracking performance Tracking efficiencies vs pt Transverse momentum resolutions dn ch /d =1000 Efficiency = Found good tracks Generated good tracks Pt/Pt = pt(measured) - pt(particle) pt(particle) dn ch /d =4000
Integrated contamination Integrated contamination Integrated tracking efficiency Integrated tracking efficiency Sequential tracking performance Magnetic field: 0.2T Magnetic field: 0.4T
Sequential tracking - secondaries Applying a second tracking pass, taking input all unassigned clusters from the first pass. No vertex constrain is imposed on the track follower (conformal mapping done with respect to the first associated cluster on track) Including all secondaries (in TPC) from K and
Sequential tracking dataflow Optical links Front-End Processors: HLT-RORC (Readout Receiver Card) Sector processors Global/event processors
Sequential tracking computing requirements CF Tr. Find. Tr. Fit. Tr. Merg. Gl. Tr. Merg. Ignoring any overhead from interprocess communication and cluster management of the HLT system Pentium III, 800 MHz Pentium 4, 2800 MHz dn ch /d CPU-time [s] #CPU CPU-time [s] #CPU 1000 7.5 1500 3.4 680 2000 14.0 2800 6.3 1260 4000 29.5 5900 13.2 2650 6000 47.3 9460 21.2 4240 #CPU = Equivalent number of required CPUs @ 200 Hz processing rate
Iterative tracking the algorithms Input: ADC-sequences above threshold Hough Transform - Apply the circle HT directly on raw ADC-data - Track candidates corresponds to local maxima in 2D parameter space Input: Track candidates Cluster Fitter - Fit the clusters along the trajectories - Track parameters provide input parameters to the fit - deconvolution is done based on the knowledge of the cluster shapes 2D gauss-fit (5 parameters) - position in pad and time - widths in pad and time (keep fixed) - amplitude
Iterative tracking - performance Tracking efficiencies Transverse momentum resolution dn ch /d =1000 dn ch /d =4000 Efficiency loss due to high number of false peaks in HT
Iterative tracking dataflow Optical links Front-End Processors: HLT-RORC (Readout Receiver Card) Sector processors Global/event processors
Iterative tracking computing requirements Hough Transform: (dn ch /d =4000) 3.5 s per sub-sector * 6 * 36 = 750 s for complete TPC event 150 000 CPUs @ 200 Hz processing rate (PIII 800 MHz) Heavy IO-bound algorithm, as extensive and repetitive access to memory is needed (filling histograms). Algorithm is however local by nature, with a high degree of parallelization. Suitable for implementation in FPGA co-processor on the HLT-RORC.
TPC data compression Applicable compression techniques Implemented compression scheme Impact on the data Expected compression factors
Applicable TPC data compression techniques Local techniques: Applied on the scale of ADC-data. ( NIM A489 (2002) 406 ) Prob. distribution of ADC-values Lossless Lossy Type of encoder Relative event size [%] Arithmetic Coding 80 Huffman Coding 71 Vector Quantization 64-48 Global techniques: Applied on the scale of clusters and tracks. Cluster position represent small deviations,, from the track fit and is subject to detector resolution. Cluster widths are a function of track parameters Lossy compression, however, relevant information is contained in the reconstructed observables. Describe the clusters within the track model, and store only deviations from the model. Track and cluster modeling
Track and cluster modeling Track parameters Size (byte) 4 (float) DCAO 4 (float) r DCAO 4 (float) Z DCAO 4 (float) 4 (float) HLT code Cluster parameters Size (bit) Cluster present 1 Sector / change 6 / 1 Total charge 12 Pad residual n( ) Time residual n( ) Offline code n( ) is a function of quantization intervals, : pad,q = pad pad n( ) = max( log( ) pad,q ) + 1 log(2)
Impact on the efficiencies Removing all remaining clusters Keeping selection of remaining clusters Loss due to parameterizing the cluster widths.
Impact on resolutions Transverse momentum resolution vs pt Integrated transverse momentum resolution vs Residuals before compression dn ch /d =1000 Residuals after compression
Results and implications RLE 8 bit TPC data Compressed track-cluster data dn ch /d Event size [MB] Data rate [MB/s] Event size [MB] Data rate [MB/s] 1000 13.8 2760 1.8 360 2000 24.1 4820 3.3 660 4000 44.1 8820 5.9 1180 6000 61.6 12320 7.8 1560 Constrained bandwidth to mass storage: 1200MB/s
Arithmetic coding of the residuals Entropy of the residual distributions suggests further (lossless) compression Fixed bitrate: 7+1 bits Fixed bitrate: 8+1 AC can theoretically achieve factor 2-3 better compression BUT: Overhead from track parameters and fixed cluster parameters (charge, headers) Overall improvement in compression ratios of ~20%
Concluding remarks and outlook Online pattern recognition: Sequential approach efficient in the lower multiplicity range, and is superior to offline with respect to computing requirements. Iterative approach does not show any improvement at higher multiplicities, which is mainly due to high number of false identified peaks in the HT Data compression: Compression factors of 6-10 is achievable. Loss is strongly dependent on the performance of the preceding pattern recognition step
Concluding remarks and outlook Online pattern recognition: Assuming current predictions on particle multiplicity, sequential tracking method is a good candidate. However, given the huge uncertainties more effort is needed to extend the tracking capabilities towards higher multiplicity: - Combine conventional tracker and iterative tracking approach Other detectors needs to be included in the framework, in particular the ITS and TRD. Data compression: Further optimization only makes sense using real TPC data. Likely not to be used during the first years of operation.