Overlap Removal and Timing Optimization Studies

Nicole Carlson, Northwestern University

8/8/07Supervisor: Tomasz Bold

2

ATLAS Trigger System

Allotted 40 ms

Allotted 4 s

3

RoI’s in the LVL1 Detector

4

HLT Selection Process

2 most Important Features:

1) Early Rejection Principle

2) Regions of Interest (RoI’s)

RoI’s from LVL1

Chain Signature

Algorithm

TE’s

5

Reason for the Overlap Algorithm

• If one wants a jet + electron in one event from the calorimeter, it is important when looking at the combined signature to make sure they come from different eta/phi

• If two TE’s have exactly the same RoI eta/phi position, then they probably originated from the same cluster

• The chain needs to be stopped (since in reality, both TE’s describe the same physical object)

6

Basic Method

• ComboAlgo to compare 2 or more types• Path:

offline/Trigger/TrigAlgorithms/TrigGenericAlgs/AlgoOverlap.h (NEW PACKAGE!)

• Compare the actual center distance between two RoI’s with a user-defined minimum center distance

• If the actual distance (of the specific TE’s) is smaller, then the chain is stopped

• Otherwise, the chain continues

7

Navigation Structure (w/o features) of 1 Event

Electron and jet overlap

so no TE created

Electron and jet overlap

so no TE createdElectron and jet overlap

so no TE created

8

Extra Features: 1)

• Higher multiplicity Triggers can be handled by this algorithm

• e.g. with jet + electron + tau, the algorithm will compare jet, e; jet, tau; e, tau (within the same event)

• Given 3 TE, if one of the combinations fails (say jet, e), then the whole set will fail and no output TE will be created

9

Extra Features: 2)

• User can specify minimum eta and phi distances instead

• This allows for tighter control along one coordinate

• Algorithm rejects those pairs which have both actual eta and phi distances smaller than the user-defined mins

10

Timing Studies 1: Overhead of timing, monitoring, and caching

• 4 Cases:

• 1) Timing & Validation Enabled

• 2) Timing Disabled/Validation Enabled

• 3) “Online Mode”: Timing & Validation Disabled

• 4) Caching Disabled, Timing & Validation Enabled

11

Results200 t-tbar events LVL2 Total Time in various modes

578.7 463.4 466

1044

0200400600800

10001200140016001800

Timing Enabled Timing Disabled Timing andValidationDisabled

CachingDisabled

Mode

Tim

e (m

s)

200 t-tbar events EF Total Time in various modes

3504 3446 3468

4977

01000200030004000500060007000

Timing Enabled Timing Disabled Timing andValidationDisabled

CachingDisabled

Mode

Tim

e (m

s)

12

Key Points

• Disabling the timers reduces consumption in LVL2 but not in EF

• Disabling monitoring does not seem to have an effect

• Caching reduces the time by approximately a factor of 2

13

Timing Studies 2: Overhead of Framework of Algorithms

• The algorithms take a certain amount of time to reconstruct physical objects

• These are run in a steering framework which schedules them so find this scheduling time

• Determine this time by replacing real algorithms with dummy algorithms which do nothing

14

Results

• Overhead of Framework:

• LVL2: ~10%• EF: ~ 4%

L2 Algorithms Overhead

35.98

331.6

0

100

200

300

400

500

600

Dummy Algorithms Real Algorithms

Test

Tim

e (m

s)

EF Algorithms Overhead

57.5

1568

0

500

1000

1500

2000

2500

Dummy Algorithms Real Algorithms

Test

Tim

e (m

s)

15

Fun Stuffs

16

Thanks

• Tomasz Bold, Till Eifert

• NSF, University of Michigan, CERN Summer Student Programme

• Jeremy Herr

• All the other summer students