High Level Trigger HLT

1
High Level Trigger (HLT)

• Status Report
• Bergen Budapest Frankfurt Heidelberg Oslo

2
High Level Trigger

• HLT tasks
• Physics case studies
• Realtime pattern recognition and (sub-)event
  reconstruction
• System architecture
• Resources
• Outlook

3
HLT tasks

• Trigger
• Accept/reject events
• verify dielectron candidates
• sharpen dimuon transverse momentum cut
• identify jets
• ...
• Select
• Select regions of interest within an event
• remove pile-up in pp
• filter out low momentum tracks
• ...
• Compress
• Reduce the amount of data required to encode the
  event as far as possible without loosing physics
  information

4
Data flow (HLT only)
TPC
detector
TRD
detector
Other detectors
data volume
30
Mbyte
/
event
Acquisition time 80µs
Acquisition time 2.0µs
front
-
end
front
-
end
front
-
end
electronics
electronics
electronics
25 Gbyte/sec
DDL
FEPLDC
FEPLDC
FEPLDC
FEPLDC
FEPLDC
FEPLDC
High Level Trigger System
lt 2 Gbyte/sec
Event Building
lt 1.25 Gbyte/sec
Permanent Storage System
5
Physics case studies

• Quarkonium spectroscopy
• Dimuons
• Dielectrons
• Open Charm
• Jets
• ...

6
Quarkonium - dimuons (1)

• Utilizing tracking chamber information
• Improving momentum resolution
• Vertex pointing capability

F. Manso et al., ALICE-INT-2002-04
7
Quarkonium - dimuons (2)

• Sharpening of pt-cut
• Preserves signal
• Rejects background
• Rejection factors
• Low pt-cut gt 4
• High pt-cut gt 100

Reduction in event rate
F. Manso et al., ALICE-INT-2002-04
8
Quarkonium - dielectrons (1)

• TRD trigger rates
• ptsingle gt 1 GeV/c ptsingle gt
  0.8 GeV/c ptpair gt 3 GeV/c
• J/?/event 0.007 0.0006
• background/event 0.39 0.15

HLT system
TRD
TPC
Online track reconstruction 1) selection of
eepairs (ROI) 2) analysis of
eepairs (event rejection)
9
Quarkonium - dielectrons (2)

• Trigger strategy
• Re-analysis of TRD tracking and PID
• Precise tracking of dielectron candidates in TPC
• Combined track fit TRD-TPC-(ITS?)
• Additional PID by dE/dx
• Rejection of background tracks (mainly
  misidentified pions) by combined (TRDTPC) PID
• Rejection factor (estimated)
• 5 (singles)
• 25 (pairs)

10
Quarkonium - dielectrons (3)

• Event flow

L0
L0
TRD
Global
TRD
Global
L0
L0
Trigger
Trigger
Trigger
Trigger
2 kHz
2 kHz
L1
L1
Readout
Readout
L1
Readout
L1
Readout
Other Trigger
Other Trigger
L1
L1
other detectors
other detectors
TPC
TPC
Detectors
,
Detectors
,
L0pretrig.
L0pretrig.
L2
accept
L2
accept
L2
L2
(
216
Links, 83 MB/
evt
)
(
216
Links, 83 MB/
evt
)
TRD
TRD
Zero
suppressed
Zero
suppressed
e

e
-
tracks
e

e
-
tracks
TPC
Data
Detector Data
Sector
parallel
Sector
parallel
Detector raw data readout
Select
Select
regions
of
seeds
Event sizes and
regions
of
seeds
Event sizes and
Event sizes and
Event sizes and
interest
interest
number of links
number of links
HLT
HLT
number of links
number of links
TPC only
TPC only
TPC only
TPC only
Tracking
of
Tracking
of
On
-
line
Data reduction
On
-
line
Data reduction

-

-
e
e
candidates
(
tracking
,
reconstruction
,
e
e
candidates
(
tracking
,
reconstruction
,
inside
TPC
inside
TPC
partial
readout
,
partial
readout
,
compression
)
compression
)
enable
enable
Verify
e

e
-
Verify
e

e
-
hypothesis
hypothesis
Track
segments
Track
segments
e

e
-
tracks
e

e
-
tracks
Time, causality
Time, causality
Reject
Reject
And
space points
And
space points
Plus
ROIs
Plus
ROIs
event
event
Binary loss less Data compression
(RLE, Huffman, LZW, etc.)
Binary loss less Data compression
(RLE, Huffman, LZW, etc.)
0.5
-
2 MB/
evt
4
-
40 MB/
evt
45
MB/
evt
0.5
-
2 MB/
4
-
40 MB/
ev
45
MB/
ev
B
C
11
Open charm

• Hadronic charm decays
• D0 ? K ?
• high pT of the decay products
• 50 of decay pions have pT gt 0.7 GeV/c

• Background
• subevent direct pions
• low- pT tracks
• 15 of charged particles have pT gt 0.7 GeV/c

12
Open charm momentum filter (1)

• Trigger strategy
• find high-pT tracks in outer sector of TPC (based
  on seeds from TRD)
• extrapolate track back to vertex
• record raw data along trajectory
• Problem of overlapping clusters
• for deconvolution of high-pT track clusters the
  knowledge of track parameters of crossing tracks
  is necessary
• Solution
• reconstruction of all tracks in the neighborhood
  (same/neighboring sector and ? )

13
Open charm momentum filter (2)

• Momentum filter
• pt gt 0.7 GeV/c
• Trigger efficiency
• signal loss lt 50
• background data volume reduction factor
  7

Offline pt-cut 0.6-0.8 GeV/c
14
Fast pattern recognition

• Sequential feature extraction
• cluster finder and track follower on space points
• low multiplicity events
• pp pileup removal
• PbPb, outer padrows
• Iterative feature extraction
• tracklet finder on raw data (Hough transform) and
  cluster evaluation
• high multiplicity events
• Simultaneous track following and track fitting
• Kalman filter
• selected tracks
• detector merging

15
Fast cluster finder and track follower (1)

• Reconstruction efficiency in pp
• ? 98
• Pileup removal
• Full event reconstruction

pp pileup event one triggered event 20 ghosts
16
Fast cluster finder and track follower (2)

• Cluster finder performance at low occupancy
  (central PbPb event, outermost padrow)

17
Fast cluster finder and track follower (3)

• Reconstruction efficiency in low multiplicity
  PbPb events cluster finder
  tracker follower

• Statistical efficiency evaluation
• of found clusters / track padrow crossing
• of good tracks / of reconstructable
  tracks
• Individual evaluation - work in progress

18
Hough transformation

• Tracking in high multiplicity PbPb events

19
Event reconstruction vs data compression

• Event rejection sub-event reconstruction
• High precision pattern recognition
• tight cuts, Kalman re-fitting
• precise high-pt tracking
• Momentum filter event reconstruction
• Medium resolution tracking
• merging of Hough tracklets, refitter
• Compression data modeling
• Redundant tracking with open cuts
• local track model for cluster collection and
  cluster unfolding

20
High Level Trigger architecture (1)
TPC sector 1
TPC sector 2
...
...
ITS
TRD
Trigger detectors Dimuon, TRD, ...
sub-sector level - 216 nodes
sector level - 108 nodes
.........
.........
supersector level - 72 nodes
event level - 12 nodes
trigger level - 1 node
(sub/super)-sector numbers TPC only
21
High Level Trigger architecture (2)

• TPC-sector readout

• Subsector node - FEP/LDC

DDL
x18 x2
FPGA co-processor
DIU
cave
processor memory PCI
counting house
6 nodes per TPC sector
...
network
network
Event building
Event reconstruction
HLT Network
22
HLT RORC

• FPGA co-processor for data intensive tasks
• Local pattern recognition on-the-fly
• Cluster finder
• Hough transformation tracker

PCI bus
DIU
-
CMC
FPGA
Memory
PCI bridge
Glue logic
interface
Coprocessor
D32
³
internal
2 MB
DIU card
SRAM
2 MB
Memory D32
23
HLT RORC
HLT RORC

• Floor plan of PCI board

• Floor plan of PCI board

24
HLT RORC

• Rubber bands

25
FPGA co-processor

• VHDL implementation of fast cluster finder

Testbench
DIU
26
FPGA co-processor

• High Level Design State machine VHDL code

VHDL code
27
FPGA co-processor

• Simulation with TPC data verification of
  functionality

28
FPGA co-processor

• Verification of functionality
• cluster finder results
• C code VHDL code

29
Interprocess communication publisher/subscriber
model
Shared data structures (not moved physically)
30
Interprocess communication publisher/subscriber
model

• Analysis processes made up of three parts
• subscriber as data input for new event data
• processing code that works on the data in shared
  memory and possibly writes some new output data
  into shared memory
• publisher to make the resulting output data
  available

Read data
Write data
31
Status publisher/subscriber

• Networks supported (status)
• TCP/IP (available) Max. Event Rate w. 128 B
  Events gt2.2 kHz
• (TCP class library used sends
  messages at gt45 kHz)
• Ethernet ST (planned)
• SCI (prototype)
• ATOLL (planned)
• Performance (on same node)
• Maximum event transaction rate gt12.5 kHz
• CPU time per complete event loop lt80 µs
• Only operating system primitives used for
  communication (e.g. named pipes)
• C source code for Linux at www.ti.uni-hd.de/HLT

Reference Platform 733 MHz Dual Pentium,
ServerWorks Chipset
32
System test

• Heidelberg cluster
• 38 Dual PIII nodes
• 733/933 MHz
• 512 MB
• Networks
• 100 MB Ethernet
• 1000 MB Ethernet
• SCI 2D Thorus
• Built-in Fault Tolerance

• Bergen cluster
• 11 Dual PIII nodes
• 833 MHz
• 512 MB
• Networks
• 100 MB Ethernet
• 1000 MB Ethernet
• SCI ring

33
Heidelberg HLT cluster
34
System benchmark

• pp pileup removal at 270 Hz (TCP/IP,
  fast ethernet)

Rate gt270 Hz
Slice Merger
SM
Event Gatherer
EG
CPU 100
Patch Merger
PM
Event Merger
EM
Netto 6131 kB/sec Brutto 6374 kB/sec
EG
Event Gatherer
EG
CPU 75-100
T
T
T
T
T
T
T
T
T
T
T
T
Tracker (2x)
ES
ES
Event Scatterer
Netto 61.8 MB/sec Brutto 62.4 MB/sec
CF
CF
Cluster Finder
CPU 60-70
AU
AU
ADC Unpacker
FP
FP
File Publisher
Patches
1
6
5
4
3
2
35
HLT team (1)

• Frankfurt
• IKF
• R. Stock
• C. Loizides (RORC)
• R. Bramm (CERN technical PhD student HLT-DAQ)

• Heidelberg
• KIP
• V. Lindenstruth
• L. Hess (fault tolerance)
• M. Kirsch (f. t.)
• S. Phillip (f. t.)
• F. Pister (f. t.)
• T. Steinbeck (publisher/subscriber)
• A. Wiebalck (cluster infrastructure)
• 2 students

• Budapest
• RMKI
• Z. Fodor

36
HLT team (2)

• UiO
• Nuclear physics group
• G. Løvhøiden
• T. Tveter
• T. Vik (simulation)
• Electronics group
• B. Skaali
• D. Wormald (RORC)
• 1 student (RORC)

• UiB
• Nuclear physics group
• D. Røhrich
• A. Vestbø (pattern recogn.)
• 2 students (pattern recogn.)
• Microelectronics group
• K. Ullaland
• 3 students (RORC)

• HiB
• Technical computing group
• H. Helstrup
• K. Fanebust (simulations)
• J. Lien (RCU/RORC)
• 10 students (Cluster, RORC)

37
Outlook

• Debugging of HLT RORC PCI board
• Synthesis of VHDL cluster finder
• System test including HLT RORCs
• Trigger efficiencies for various applications
• Optimization of Hough transformation tracker
• VHDL implementation of Hough transform
• HLT-DAQ interface
• Conceptual design review

38
Additional slides
39
Data compressionEntropy coder
Probability distribution of 8-bit TPC data

• Variable Length Coding
• short codes for long codes for
• frequent values infrequent values
• Result compressed event size 72

40
Data compressionVector quantization

• Sequence of ADC-values on a pad vector

• Vector quantization transformation of
  vectors into codebook entries
• Quantization error

compare
code book
Result compressed event size 29
41
Data compression TPC-data modeling

• Fast local pattern recognition
  

• Track and cluster modeling

simple local track model (e.g. helix)
local track parameters
track parameters
comparison to raw data
analytical cluster model
quantization of deviations from track and
cluster model
Result compressed event size 7
42
HLT RORC

• Floor plan of PCI board

43
FPGA coprocessorImplementation of Hough
transform
44
Network performance

• TCP/IP over Fast Ethernet (100 Mbit/sec)