Track quality - impact on hardware of different strategies

16
Track quality - impact on hardware of different strategies Paola FTK meeting 21-02-2007 1. Performances on WH and Bs 2. Now we use all the layers in the fit and we suffer– can we drop something going to raw hits? 3. Refresh about the FTK architecture 4. Impact on hardware due to different configurations

description

Track quality - impact on hardware of different strategies. Performances on WH and Bs  mm Now we use all the layers in the fit and we suffer – can we drop something going to raw hits? Refresh about the FTK architecture Impact on hardware due to different configurations. - PowerPoint PPT Presentation

Transcript of Track quality - impact on hardware of different strategies

Page 1: Track quality - impact on hardware of different strategies

Track quality - impact on hardware of different

strategiesPaola

FTK meeting 21-02-2007

1. Performances on WH and Bs 2. Now we use all the layers in the fit and we suffer– can

we drop something going to raw hits?3. Refresh about the FTK architecture 4. Impact on hardware due to different configurations

Page 2: Track quality - impact on hardware of different strategies

FTK iPat

iPatV10

Truth

Page 3: Track quality - impact on hardware of different strategies

3

FakesBs

GoodBs

89%

10%

5/6 tracks2 pixel layers5/6 tracks

3 pixel layers

Page 4: Track quality - impact on hardware of different strategies

23/01/2007 Francesco Crescioli 4

Bs-> fake analisys - cuts

Asym in ctheta distance.To be investigated.

Distance from the nearest truth.

Log scale.

C phi

cthetaz0

Page 5: Track quality - impact on hardware of different strategies

5

Nomiss

MissPix0

MissPix2

MissSct0

MissSct1

MissSct2

MissSct3

2 pixel layers + 4 SCTSpace points

3 pixel layers + 3 SCTSpace points

Nomiss

MissPix0

MissPix2

MissSct0

MissSct1

MissSct2

MissSct3

Page 6: Track quality - impact on hardware of different strategies

6

Were is the problem? Francesco can you run your classification fake/good on WH samples, or give to Monica the tool? It could be “higher fraction of Fakes”

1. Is it in FTK or is it a more general problem? What happens if FTK and iPatRec use exactly the same detector? NO TRT, all the 7 available space points for both.

2. Is it the new software release that has problems?3. Can a complex b-tag algorithm recover some performances?

How today-iPatRec b-tag performance compare with the past?

4. If the problem is the poor detector performance (problem seen both by iPatRec & FTK), difficult to use only 6 layers in FTK dropping redundance: is it a pattern recognition problem or a fit problem? Just a fit problem: we can use 11 layers in the fit and 6

layers in the AM General problem: may be we need to use 11 layers in the

whole system since the prototype beginning.

Page 7: Track quality - impact on hardware of different strategies

7

HARDWARE ARCHITECTURE

Pixels barrel SCT barrel Pixels disks

Pt (GeV)

Rcrv (m)

(deg)

(rad)

2 3.333 4.5 0.0781

1 1.666 9.0 0.1567

0.5 0.833 18.2 0.3071

Size of overlap regions as function of Pt threshold

AM

R-stereo: 2 SCT layers/bus

6 busesEach oneCan provide 2 layers

1 core crate/partition

Page 8: Track quality - impact on hardware of different strategies

8

Track dataROB

Track dataROB

Raw dataROBs

~Offline quality Track parameters

~75 9U VME boards – 4 types

SUPER BINSDATA

ORGANIZERROADS

ROADS + HITS

EVENT # N

PIPELINED AM

HITS(LVDS links)

DO-board

EVENT # 1AM

-board

2nd step: clean-up & track fitting

Pixels & SCT

DataFormatter

(DF)

50~100 KHzevent rate

RODsRODs

cluster finding?split by layer

overlap regions

NEW

S-links

RW-HWTrack Fitter

Page 9: Track quality - impact on hardware of different strategies

9

CORE system SIZE

CUSTOM BACKPLANE

FTK INPUTFROM

Data Formats

O(2.5 M) patterns 1 crate forEach sector

AM

-B7

AM

-B8

AM

-B1

AM

-B0

DO

5D

O4

DO

3D

O2

DO

1D

O0

Gh

ost

B

ust

er

CPU

0 C

PU

1 AM

-B2

AM

-B3

CPU

2 C

PU

3 AM

-B4

AM

-B5

AM

-B6

1 Gigafitter

128 AMChips/board

= 320 kpat/boardUsing 12 layers640 kpat/boardUsing 6 layers

NOW in FTKsimSS size (too large? Check timing!):Pix: 5mmx z_ModSct:10mmx z_Mod~1 Mpat for90 degree sector:

4x2 AMB for 12

layers2 AMB for 6 layers

8x2 AMB for 12 layers

4 AMB for 6 layers

180 degree sector:

Page 10: Track quality - impact on hardware of different strategies

10

SCTend caps

x3 wheels

x3 wheels

z

Pixel end caps

6 modules each

8 sectors option

DATA FORMATTERs

Barrel

Page 11: Track quality - impact on hardware of different strategies

S-links

S-links

Receive S-linksFind ClustersDivide by LayersSend Hits to DOsSends overlap region to near phi-sector

DATA FORMATTERLVDSSerializer/Deserializer

To DOs

Overlapregion

P3

DATA FORMATTER

Page 12: Track quality - impact on hardware of different strategies

12

RODs and DFs if RO by wedge

# RODs

# DFs4 S-links/DF

# DFs6 S-links/DF

# outputs

/DF

Pix 0 36 9*->10 6 1 (1 layer)

Pix 1-2 32 8 5*->6 2 or 1 if Muxed

Pix disks 16 4 3*->4 5

SC0-3 44 11*->12 8 4 (8 stereo)

SC disks 48 12 12 3 (6 stereo)

Total 176 46 24+12=36 130/106

180 degree sectors

Page 13: Track quality - impact on hardware of different strategies

13

RODs and DFs if RO by wedge

# RODs

# DFs4 S-links/DF

# DFs6 S-links/DF

# outputs

/DF

Pix 0 36 10 6 1 (1 layer)

Pix 1-2 32 8 6 2 (1 if Muxed)

Pix disks 16 4->8? 4->8? 5

SC0-3 44 12->16 8 4 (8 stereo)

SC disks 48 12->16 12->16 3 (6 stereo)

Total 176 46->58 36->44 130/106

45 degree sectors

If we have 8 sectors some DF needs to send data to 2 different DOs.This is feasible for DF with few outputs (Pix0, Pix1-2, Pix disks?).It can be done for SC0-3 and SC disks too if stereo layers are not used or sent to the same DO.The fall back solution is to increase the number of DFs up to nearest multiple of 8.

Page 14: Track quality - impact on hardware of different strategies

14D

O5

DO

4D

O3

DO

2D

O1

DO

0G

host

Bust

er C

PU

0 C

PU

1 C

PU

2 C

PU

3 AM

-B7

AM

-B8

AM

-B1

AM

-B0

AM

-B2

AM

-B3

AM

-B4

AM

-B5

AM

-B6

SLINKS

DO

5D

O4

DO

3D

O2

DO

1D

O0

Ghost

Bust

er C

PU

0 C

PU

1 C

PU

2 C

PU

3 AM

-B7

AM

-B8

AM

-B1

AM

-B0

AM

-B2

AM

-B3

AM

-B4

AM

-B5

AM

-B6

RO

Ds

DF

DF

DF

To DOs

LVDS LINKsThe Old IDEAs: 2 phi sectors>16 AMB for 12 layers: too much

Overlap

Page 15: Track quality - impact on hardware of different strategies

15

Let’s try to work with 4 sectors. First of all try to see if a BANK with 11 LAYERS is REASONABLE. We expect a factor 2 AM size

AM

4 core crates, each one containing:• > 10 AMBoards• 6 Dos• 1-4 boards for RW-HW-Track fittingDFs # ~ 36 independent on sector # ~ 2 crates. TOTAL 6 crates

If we would Use 2 trackSegments: If 2 sectors 4+2 cratesIf 4 sectors 8 + 2 cratesBut worse performances

Page 16: Track quality - impact on hardware of different strategies

16

Conclusions

1. First of all understand well the problems in track reconstruction quality. Repeat Francesco analysis of WH sample and see if a good b-tagging algorithm can make the difference

2. However I suspect that going to raw hits we cannot use less than 11 layers at least in the fit (differently from what we were planning for the first prototype) !

3. Let’s try to use a 11 layer for pattern recognition and fit:• Let’s see how much big is the bank (we expect x 2)• Let’s see the performance of the system (road size and

processing time also!)

4. If we find problems with an 11 layer AM bank :• Use 6 layers in AM and 11 layers in fit procedure• Or see if we can split into 2 track segments.