physics data processing - the online connection

Physics Data Processing -The online connection

Nikhef colloquium9/2/2009

Sander Klous

Physics Data Processing - Sander Klous 2

Bob Hertzberger

The steamboat view

PDP Activities9/2/2009

My focus

Remote O

nline Farms

Use

Cas

es (i

.e. t

rigge

r per

form

ance

)


Remote Online Farms

• What is a remote online farm?• The connection to the online system• Event routing and streaming• Remote online farms in ATLAS (Hegoi Garitaonandia)

– Debug stream reprocessing– Trigger menu validation– Muon calibration stream

• Do we really need remote online farms?9/2/2009


Concept

9/2/2009

Amsterdam

NIKHEF/SARA

Data Acquisition40 MHz

Level 1

Level 2

Accept 1 in 500

Accept 1 in 50

Accept 1 in 10

Level 3

Computing grid

Networkswitch


Why is this interesting?

9/2/2009

Gary Stix, editor of Scientific American

Janu

ary

2001


Connection to the online system

9/2/2009

ROB

Data Collection Network

L2PU

EFEvent Filter

Back End Network

SFOMassstorage Building 513

ROF

Remote Event Processing Farms

ROF

ROFROF

Packet

Switched

(GEANT)

Ligh

t pat

hLocal Farm

ROBROBROB

SFISFISFILevel 2 Trigger

DDM

Routing and Streaming

Catalin Meirosu


Routing and Streaming introduction

9/2/2009

• Optimizing resources– Online, different routes– Offline, different streams

• Classification of events– Physics– Calibration– Debug– Express– Remote

Hans von der Schmitt


Regions of Interest andPartial Event Building

9/2/2009

Ignacio Aracena

Virtual Point 1


Inclusive or Exclusive streaming

9/2/2009

Hans von der Schmitt


Online monitoring of overlaps

9/2/2009

Brian Petersen


Luminosity blocks

9/2/2009 Hans von der Schmitt


raid25TB

TAGDB

load TAG

recon EXPR

recon PHYS

calib/align

disk150TB

CondDB

Dataacquisition

TIER-0

CASTOR

CPU farm

DDM/DQ2

TIER-1s

3201060

130

440

720

340

720

320

200

200

320

200

200

online

c1

c1

RAW

ESD

AOD

TAG

Express stream handling

9/2/2009

T0MDB

SFO

Rates in MB/s

Final Dress RehearsalData Quality Monitoring

Luc Goossens


Debug stream reprocessing

9/2/2009


Trigger menu validation

9/2/2009


Muon calibration

9/2/2009

Virtual Point 1

MuonCalibration

Group


Do we really need ROFs?

• At the moment: No, not really…– Not for physics at least

• Transition will be gradual– Example: debug stream reprocessing– Limited resources in the CERN Analysis Facility

• Funding will play an important role– Easier to fund online resources in home country– Other factors: energy consumption, human resources

• What will be the first physics use case?9/2/2009


Trigger performance

• Fully Hadronic decays of top pairs• The jet trigger challenge

– Determining trigger efficiencies from data– Tag and probe with leptons– T&P with semi-leptonic decays of top pairs

(Menelaos Tsiakiris)

– Turning the probe around, T&P with jets– Jet trigger efficiencies and rates

• Beyond multi-jet triggers– Topology triggers and remote online farms

9/2/2009


Fully Hadronic decaysof top-antitop pairs

9/2/2009


The jet trigger challenge

• QCD 6 jet background– 1000 to 10000 x Signal

• Is this possible at all?– Long term study

• First look at semi-leptonic decays– Study the possibility to trigger on jets– Extract jet trigger efficiency from data– Cross correlate with muon triggers9/2/2009

Turn on kicks in.Trigger problem

LambacherMunichO(1 fb-1)


Tagged Tagged and Probed

Trigger efficiency from data

Tag and probe• A first example:

di-muon decays of Z bosons

9/2/2009

20 GeV muon turn on curve (CSC note)


Semi leptonic decays of top pairs

• Tag on the jet side and probe on the muon side• Extract number of signal events (1 fb-1)

– Invariant mass reconstruction on hadronic side

9/2/2009

Possible problem: contamination with tau to muon decays


Event by event trigger efficiency

• Determine individual object efficiencies– Extrapolate T&P from di-muon decays of Z-bosons

• Combine objects into an event efficiency• Weigh the event appropriately

9/2/2009

Correct isolation withrespect to closest jet

Correct foreta dependencyWithout corrections

Tamsett – Royal Holloway


Turning the probe around

• More complicated– The jets are multi-object triggers,

e.g. 4J_95: 4 jets above 95 GeV– Thresholds are likely to change

• The missing link:– Trigger efficiency of Semi Leptonic decays– Fully Hadronic decays of top quark pairs

• Is it possible?– If not, can we do better?

9/2/2009


Topology triggers and remote online farms

Topology triggers• Time consuming

– Invariant masses– Likelihood fits

• Budget in point 1– 1500 machines x 8 cores per machine = 12000 cores– 3 kHz EF input rate from LVL2– 12000 / 3000 = 4 seconds per event (all triggers)9/2/2009

Xin Wu

Event FilterMean: 1.57 seconds

2

2

2

2

2

2

2

2

244

12

22

2.3

)175(

2.4

)(

2.4

)4.80(

3.9

)4.80(

01.0

1)(

hadTop

hadTop

lepTop

hadW

lepW

i i

i i

pcorT

pcorTi

MMMMM

EE

(Erik van der Kraaij)


Conclusion

• Remote Online Farms– An interesting research topic– Several farms already in production– Still far from enabling grid for online processing

• Physics Use Cases– Most likely complicated topologies– Fully hadronic decays of top quark pairs– Studies ongoing to understand trigger efficiencies

• The future of grid is in the online connection9/2/2009

physics data processing - the online connection

Career

online resources

jet trigger challengeqcd

remote online farmswhat

jetsjet trigger efficiencies

streamingremote online

online systemevent routing

event efficiencyweigh

jet background1000