Vertex 2001 Brunnen CH

1
Talk Summary

• Will describe CDF Silicon Vertex Tracker (SVT)
• SVT is part of CDF Level 2 Trigger
• Design
• how it works
• what is special about it
• Construction
• how much hardware, which choices, installation
• Performance
• results from real Tevatron data in 2000 and 2001
• Commissioning experience
• what would we change if we could go back
• Conclusions

2
Impact Parameter
3
The SVT Algorithm (I)Pattern Recognition

• First stage at coarser resolution
• Hardware implemented pattern recognition at SVX
  readout speed (similar to lookup table
  approaches, almost no extra-time)
• 32.000 roads for each 300 slice
• 250 micron Superstrips
• gt95 coverage for Ptgt2GeV

4
The SVT Algorithm (II)Track Fitting
5
The Device (II)
AM Sequencer
SuperStrip
AM Board
Hit Finder
Hits
Detector Data
Matching Patterns
Roads
Roads Corresponding Hits
Hit Buffer
L2 CPU
Tracks Corresponding Hits
Track Fitter
6
SVT for Real!
7
SVT Completed
8
CDF Trigger Room
SVT
9
Once it is builtthings we could not simulate

• The hardware must work correctly
• data integrity challenge
• Fitting tracks in 2-d only must be good enough
• Z-alignment challenge

10
The it works ChallengeHow to make sure the
trigger takes the right decision

• Very complex system (gt100 boards)
• Large data flow (50 cable data path at 600Mb/sec)
• Large data reduction (only 1/1000 events makes to
  tape)
• Cant feed real data without real detector and
  beam
• SVT as an event builder synchronization !
• 12 wedges x 3 Hit Finder boards x 4 G-Links
  288 optical links from SVX
• Another 24 from Layer00
• 1 stream from XTRP (XFT) fanned into 12
• L1 trigger information
• combine at 50KHz in one output stream
• One board fails ? SVT halts

11
Data Diagnostic Tools

• Simple, uniform, data-driven data paths
• Point-to-Point LVDS with asynchronous clocks and
  FIFOs
• Built-in Logic State Analyzers (Spy Buffers) on
  each end
• Allow to isolate problems to one board
• Allow data monitoring asynch to DAQ (SpyMonitor)
• Check data path synchronization at each
  data-merging point
• 8-bit BunchId tells (on average) data are coming
  from the same event
• Check data path parity and flag any illegal data
• Error flags in each board/event, freeze Spy
  Buffers on error (global or local freeze) to look
  at error cause, test signal on front panel to
  trigger external test equipment on error
  condition

12
SVT error test point used by L1 groupfrom CDF
Electronic Logbook
2
4
From SVT crate
1
..through the floor..
to XTRP rack
to find the origin of (till then) mysterious
data synchronization problem !
3
13
Z-alignment Challenge

• SVT integrates d distribution in z
• One average beam position in each 15cm SVX barrel
• Average is found and corrected for online
• Z-tilt gives resolution spread. Need lt 100?rad
• DETECTOR SVX ladders must be parallel to each
  other
• Defines an ideal line the detector axis
• BEAM must be parallel to the same line
• gross effects (once for all) corrected by SVX
  move
• fine tuning by accelerator in real time (even
  during store) needs feedback from SVT

14
Commissioning Run It Works

• October 2000
• started as SVT hardware/data-flow test
• SVX barrel 4
• Only 2 phi slices, only 2 z-barrels (instead of
  6)
• lower grade detectors, poor alignment
• No COT tracks to SVT (missing XTRP)
• Prove algorithm with simulation
• Nominal geometry
• Wide roads (500microns vs 250 design)

15
Commissioning Run (III)COTSVX II Simulated
Offline (Correcting for Beam Z-tilt)
Impact Parameter Distribution from SVT simulation
on SVX raw data and XFT output
What we promised. From SVT TDR (96) using
offline silicon hits and offline CTC tracks
s 47 mm
s 45 mm
16
2001 CDF Run It Does the Job

• SVX inserted a piece at a time
• COT tracks still suffering from problem in track
  distribution hardware (XTRP)
• No L2 processors yet
• Work focused on SVT integration and configuration
  software
• demonstration of performance
• Still using
• Wide roads
• Nominal geometry
• No SVX pedestal adjustments
• No noisy/hot/dead channel corrections
• No correction for beam offset (Dy3mm Dx1mm)

17
Our Inputs 1 - SVX

• good S/N
• resolution 15microns
• layers (mis)alignment 20microns (will correct
  for in SVT)
• 1/3 extra hit/mm/event (SVT Superstrip
  500?250mm)
• lt2.5 strips/cluster (3 used by SVT)
• had to deal with one-piece-at-a-time DAQ
  integration and data integrity problems (just
  takes time)

18
Our Inputs 2 - XFT

• XFT working well since commissioning run
• XTRP still flaky
• data integrity problem, not quality
• XFT vs. offline COT reconstruction
• D? 4 mRad
• spec was lt8mRad
• DPt/Pt 1.6
• spec was lt2

19
SVT Performance 2001 (I)Online Data d vs ? SVX
Only
28 Aug 2001 data, ?2lt40 no Pt cut
SVX only !

• Sinusoidal shape is the effect of beam
  displacement from origin of nominal coordinates
• Can find the beam consistently in all wedges
  even using only SVX

20
SVT Performance 2001 (II)Online Data d vs ?
SVXXFT
SVX COT !

• SVT real time monitor from SVT data logged in
  event record (rather than sent to L2)
• Beam position computed online in real time on
  crate controller and subtracted online in SVT
  Track Fitter
• Original plan was to do it in L2 processors

COT HV problem
21
SVT Performance (III)Fitting beam spot in each z
barrel
SVX COT

• From SVT output logged in event record (limited
  statistic)
• Still can not get data from all SVX
• Can clearly see the beam in each z-barrel to use
  for z-slope correction

22
SVT Performance (V)Dealing With Beam Z-tilt
Online

• From beam fitting program running on VME crate
  controller using Spy Buffer data (all L1 triggers
  at 50KHz) in real time
• No correction for barrel-to-barrel misalignments
• Geometry still be be understood

Beam center in the 6 barrels (15cm each)
Beam center (y)
50 ?m
200 ?m / 30 cm 600 ?rad

• Can find beam position online on each barrel
• Will feed to accelerator for real time beam
  alignment

Beam center (x)
23
SVT Performance (VI)Real Data On-line Impact
Parameter Distribution
July 8 data, each wedge fitted separately, ?2lt50
Ptgt2GeV
SVX COT

• No detector alignment correction yet
• No pedestals, noise, cluster tuning
• Not corrected for spread due to beam z-tilt
• chi2 cut can and will be made tighter
• Already almost as good as simulation on real
  Run1 data (goal is 40mm )

s 52 mm
Trigger cut
24
Experience Summary

• Very challenging project
• At the boundary of hw capabilities when started
  (96)
• Todays chips to do it right
• Debugging/commissioning very long and not
  finished
• Slowed by
• Manpower on site
• Other parts of the system being still in
  debug/commission
• Need to look at data
• Software always much more effort then originally
  thought
• Have not started to look at
• low rate problems
• Noise effect/reduction
• Pattern/fit tuning, clustering optimization

25
What We Learnt the Bad (I)things we should have
done differently

• Designed mainly for steady state operation,
  more functions and flexibility for board/system
  testing added on the road
• Plan for long, demanding, integration/commissionin
  g
• No dip-switches, ISP everywhere
• Some obsolete/inadequate hardware, overcrowded
  PLD/FPGA
• redesign/rebuild
• Did not try hard enough to make all boards more
  similar
• generic boards with different firmware ?
• Plan for lack of input/output (done, but not
  enough)
• Must be able to test SVT in place, with proper
  timing and data flow, even without
  SVX/COT/L2/beam
• Not only hw test, also operating/monitoring sw

26
What We Learnt the Bad (II)things we should
have done

• Bring to each board DAQ control signals (Clock,
  Accept, Reject, Buffer number)
• Data driven good for data flow, synchronized
  snapshots for debug
• Be able to read some internal SVT data into the
  event
• Allow surviving errors from other systems (bad
  BunchId)
• Design for expansion
• Wider data path (connector limited!)
• More SVX layers, stereo data (?)
• Some electrical details
• Do not use width-expansion in Fifo
• Clockenable instead of asynch strobe
• R/W by VME each ram/fifo/pipeline register

27
What We Learnt the Bad (III)things we did not
manage to do hard enough

• Should have invested much more in software much
  sooner
• Things we knew we had to do, planned to do, tried
  to do, but failed to do in time
• Common low level board access methods
• Common framework for board test
• Non-expert board test suites
• Bookkeeping and DB storage of configurations
• Online monitor
• Offline software
• Things we overlooked (underestimated)
• Too many things are known to too few people

28
What We Learnt the Good (I)things we are glad
we did

• Independency from CDF DAQ (data driven device)
• design/build while DAQ specs are in flux
• Test boards away from Trigger Room
• Common data communication protocol (hw/sw)
• Boards as building bricks that can be combined at
  will
• Can adapt SVT configuration to various test
  needs, both for us and others (L2 e.g.)
• Ability to inject/read data from every board
• Can test most board functions with no additional
  hardware
• Low level software captures board objects (ram,
  regs) in common framework, same high level code
  on several boards
• svtvme_sendData (board, data)
  svtvme_readFifo(fifo, data)

29
What We Learnt the Good (II)things that went
well

• Hardware fears that did not materialize
• BGA (easier then thought at least in U.S.)
• Chip replacement (done very little of)
• Better build more spare boards (cheaper)
• Paranoia about data integrity and data stream
  synchronization pays off
• Beam position finding in real time works
• A small group managed to build, operate and
  properly configure SVT
• a lot of work still to be done
• pattern/fit constants maybe easier then expected

30
Conclusion

• SVT is living up to expectations
• Track fitting with offline quality in real time
  is a reality
• Trigger rates still be measured (designed on Run1
  simulation)
• Alignment
• Narrower roads
• More material then expected
• Is an Experiment
• SVT arrived in time but
• It would have been better to build with today
  technology
• There are too many internal differences
• Would have liked better software yesterday
• Plan for commissioning and debugging, not only
  data taking

31
The Best Advice
Hike the mountains !!
32
SVT Requirements
Detector Raw Data
7.6 MHz Crossing rate 132 ns clock

• Good IP resolution
• ASAP (?10 ?sec)
• No Dead Time

Level 1 storage pipeline 42 clock cycles
Level 1 Trigger

• Level 1
• 7.6 MHz Synchromous Pipeline
• 5544 ns Latency
• 50 KHz accept rate

L1 Accept
Level 2 buffer 4 events
Level 2 Trigger

• Level 2
• Asynchromous 2 Stage Pipeline
• 20 ?s Latency
• 300 Hz accept rate

• Drop stereo info 2D tracking
• NEED BEAM PARALLEL TO SVX !!!!
• Parallel processing (12 300 slices)
• Data driven pipeline

L2 Accept
DAQ buffers
L3 Farm
Mass Storage (50100 Hz)
33
The Device (I)
The AM chip is the physical realization of the
template matching pattern recognition algorithm
each AM Chip can compare each hit with all the
patterns in memory in parallel, providing the
high speed necessary for trigger
applications. With the AM, pattern recognition is
complete as soon as the last hit of an event is
read! It is the ONLY CUSTOM PART, everything
else is RAM, FIFO, PLD, FPGA

• Full custom VLSI chip
• 0.7?m technology
• 35mm2
• 180000 transistors
• 128 patterns, 6x12bit words each
• majority logic
• Working up to 40MHz

34
An SVT Slice
35
Commissioning Run (I)Results (SVX II Standalone
Tracking)
impact parameter distribution
geometrical constraint
s 87 mm
36
Commissioning Run (II)SVX Standalone Vs Adding
COT Offline in SVT Simulation
SVT on silicon only good enough to find good
tracksloop on all SVT-COT track pairs and
compare parameters
? SVT COT
Curvature SVT - COT
37
SVT performance (IV)beam tilt is there !
Beam slope vs. run number from offline
reconstruction Unofficial average (my eye
fit) 600 ?rad x vs z 150 ?rad y vs z
38
SVT Performance (VII)Understanding of Impact
Parameter Resolution

• Best detector performance 45 micron
• from Comm.Run data offline analysis
• single wedge/barrel
• corrected for beam tilt and detector alignment
• Detector misalignment 50 micron
• from Comm.Run. analysis
• Beam Tilt 54 micron
• from current data
• Wedge-to-wedge misalignment 67 micron
• from current data
• Best width without XFT 67 micron
• Best width in present data (no corrections) 52
  micron