26
LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS and/or MCMS? Part II: Effect of 6-bit ADC Part III: Zero suppression Part IV: Data Output Rate

LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Embed Size (px)

DESCRIPTION

Planned Changes for Study Algorithms currently implemented on the TELL1 board for VELO will move to the SALT chip Reduce the ADC count from 10 bits to 6 bits Use a Common Mode Suppression (CMS) algorithm that uses all 128 channels in a chip Options are: 1)MCMS, Mean Common Mode Suppression, find the average of a set of channels, and then subtracts this from the individual readings 2)LCMS, Linear Common Mode Suppression, fit a line to a set of 128 readings and then subtract this line from the data To simplify the electronics, perform only Zero Suppression and no Clustering

Citation preview

Page 1: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms

Giulio Forcolin, Abdul Afandi,Chris Parkes, Tomasz Szumlak*

* AGH-Krakow

• Part I: LCMS and/or MCMS?• Part II: Effect of 6-bit ADC• Part III: Zero suppression• Part IV: Data Output Rate

Page 2: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Current VELO Data Processing

• Data digitized and processed on TELL1 board• 10 bit ADC• Processing algorithms currently used

– Pedestal Subtraction– Common Mode Suppression (in units of 32 chan.)– Strip Reordering– Zero-suppression and Clusterisation

Page 3: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Planned Changes for Study• Algorithms currently implemented on the TELL1

board for VELO will move to the SALT chip• Reduce the ADC count from 10 bits to 6 bits• Use a Common Mode Suppression (CMS)

algorithm that uses all 128 channels in a chip• Options are:

1) MCMS, Mean Common Mode Suppression, find the average of a set of channels, and then subtracts this from the individual readings

2) LCMS, Linear Common Mode Suppression, fit a line to a set of 128 readings and then subtract this line from the data

• To simplify the electronics, perform only Zero Suppression and no Clustering

Page 4: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Method• Used algorithms based on the Vetra algorithms• Performed algorithms in the order Pedestal

Subtraction, MCMS, LCMS and Zero Suppression

• Used integers to simulate the limited number of bits available on the electronics; this was achieved by truncating all the numbers after the units.– (i.e. equivalent to bit shifting rather than rounding)

Data SamplesNo beam data from 2012 (Run 113600)Beam collision data from mid/late 2011 (Run 98228, μav= 1.39)

Page 5: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Part I: LCMS and/or MCMS?

• Studies with 10-bit ADC

• Using data taken with no beam

Page 6: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Raw Data

• For the first step, data with no hits was studied

Page 7: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Pedestal Subtraction• Pedestals calculated as

the average of the readings of each channel

• Then subtracted from the data

• Due to integerization, the distribution was not centred around 0

• Fix by adding 0.5 to the average pedestal value, and 0.5 t during calculation,

Page 8: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Noise after MCMS and LCMS

• MCMS has a significant effect, simple algorithm • LCMS is not a great improvement, (5% improvement in noise)

• It is a lot more processor intensive than MCMS• Conclude – current data suggests only need to perform MCMS, and future proof should CM

increase• (caveat: new VELO will move closer to beam, but no sign of significant beam pickup )

Page 9: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Part II: Effect of 6-bit ADC

• Rescaling from 10-bit ADC to 6-bit ADC• But making better use of dynamic range

• Using no beam data

Page 10: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Data Scaling• Data was scaled to make it

compatible with limiting ADC counts to 6 bits

• The aim was to allow a signal of two minimum ionising particle (equivalent to ~80 ADC counts above the noise in the current data)

• [Achieved by subtracting 480 counts from the raw data, and dividing what was left by 3]

• Algorithms were then carried out as before (with truncation to 6 bits)

Page 11: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Results with 6-bit ADC• Use of 6-bit ADC increases

effective noise by 35%• Can counteract to measure

noise by rescaling ADC

10 bit processing noise = 2.38 [10 bit ADC]6 bit processing noise = 1.13 [6 bit ADC]

= 3.39 [10 bit ADC]

10 bit processing noise = 1.99 [10 bit ADC]6 bit processing noise = 0.88 [6 bit ADC]

=2.64 [10 bit ADC]

Page 12: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Results with other rescaling

• Rescaling the data by a factor of 2 increases effective noise by 15%, however it would not be possible read a 2 MIP signal using this scaling, would need 7-bit ADC.

• Using 8-bit ADC would remove the need for scaling and still save 2 bits

10 bit processing noise = 2.38 [10 bit ADC]6 bit processing noise = 1.63 [6 bit ADC]

= 3.26 [10 bit ADC]

10 bit processing noise = 1.99 [10 bit ADC]6 bit processing noise = 1.14 [6 bit ADC]

=2.28 [10 bit ADC]

[Achieved by subtracting 480 counts from the raw data, and dividing what was left by 2]

Page 13: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Part III: Zero suppression

• Using one simple threshold to zero-suppress

• Using late 2011 data (after radiation damage)

Page 14: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Zero Suppression (ZS)• To simplify the design of the electronics it has been

decided that they (if possible) should only perform Zero Suppression– Clustering would then be performed on TELL40

• This means that no reordering needs to be performed at this stage – i.e. even were it needed it could be done on the FPGA

• Zero Suppression performed by only reading out signals above a certain threshold

• E.g. look for reading 4σ above the noise level

Page 15: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Zero Suppression (ZS)

• Setting threshold of 4σ excludes almost all of the data for the no-beam data set

• Integerization causes there to be a difference in the number of readings between the two sets of data

Page 16: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Zero Suppression (ZS)• Threshold set at 10 [10-bit

ADC counts] or 3 [6-bit ADC counts]

• It appears a lot of “noise” is outputted even if the threshold is 4σ above noise level

• This “noise” is not present in no beam data, probably due to radiation damage second metal effect

• In 6-bit ADC integerization is significant (10 in 10-bit would be 3.3 in 6-bit) and can lead to significant increase in noise

6-bit ADC

10-bit ADC

Page 17: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Chris Parkes

17

VELO R Sensor

Page 18: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Chris Parkes

18

VELO R Sensor 2nd Metal Layers

Page 19: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Cluster Finding Efficiency

Chris Parkes

19

Jon Harrison

2nd Metal Layer Charge Loss

Page 20: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Part IV: Data Output Rate

• Studies of data size given different protocols for data output

Page 21: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Data output rate• Have 27-28 bits that are sent out once for each event • Channel ID and ADC count sent out for each channel

with a reading above the threshold• Number of channels that can output data limited by

payload size (above this saturate)

Data sent out once for each event Sent out for each strip with a reading above

threshold

Bunch ID Sensor ID Chip ID Payload Size

Channel ID ADC count

12 bits 7 bits 5 bits 3-4 bits 7 bits 6 bits

Page 22: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Data output Rate3 Options to output data

1) Output Bunch ID, Sensor ID and Chip ID for each bunch crossing while the Payload Size, is sent out only when at least 1 hit is present

2) Output Bunch ID, Sensor ID, Chip ID and Payload Size for each bunch crossing

3) Output Bunch ID, Sensor ID, Chip ID and Payload Size only when at least 1 hit is present

For each of these options can have a payload size of either 3 or 4 bits;therefore the number of readings that can be outputted is limited to

1) 8 or 162) 7 or 153) 8 or 16

Page 23: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Data output Rate• Assume a bunch crossing every 25 ns• Assume that there is a certain occupancy rate,

P, for each channel to register a hit1) For a probability of P = 1%, expect 5.12x107

readings per chip per second 2) For a probability of P = 2%, expect 1.02x108

readings per chip per second • Use a Poisson distribution to calculate the

probability that each chip records a certain number of hits

• Can therefore calculate the expected data output rate for each chip

Page 24: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Data Output Rate1% Occupancy 2% Occupancy

Option Payload Size (bits)

Data Output Rate (GByte s-1)

Readings Missed (%)

Data Output Rate (GByte s-1)

Readings Missed (%)

1 3 0.199 7.2x10-4 0.280 0.069

4 0.203 4.1x10-12 0.284 7.4x10-8

2 3 0.203 5.2x10-3 0.280 0.26

4 0.208 6.2x10-11 0.285 6.4x10-7

3 3 0.168 7.2x10-4 0.271 0.069

4 0.172 4.1x10-12 0.275 7.4x10-8

For a single chip

Page 25: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Observations• If payload size is limited to 3 bits, expect to lose up

to 0.26% of readings if the P ≤ 2%, however the data rate is also reduced by ~1.5% compared with 4 bits

• For P = 1%, the data rate is reduced by ~15% when sending headers only when data is present (option 3), however for P = 2%, it is only reduced by ~3%

• Option 2 increases data rate by ~2%, however should be easier to implement.

• Conclude:– 1) Need detailed knowledge of occupancy rate to make decision on

whether send headers out is significant– 2) 3-bit for payload size seems sufficient for 2% occupancy

Page 26: LHCb VELO Upgrade Strip Chip Option: Data Processing Algorithms Giulio Forcolin, Abdul Afandi, Chris Parkes, Tomasz Szumlak* * AGH-Krakow Part I: LCMS

Conclusions• MCMS needed, but not LCMS

• 6-bit ADC, integerization causes 25% increase in measured noise

• Zero-suppression in 6-bits– Needs further study. 2nd metal layer effect could cause a

problem if at upgrade.

• Data rates: 3 bit payload size, rate at 2% occupancy ~ 0.27Gbytes s-1 per chip