1. CHEP 2003 @ La Jolla San Diego 24-28/3/2003
The Algorithm Steering and
Trigger Decision mechanism of
the ATLAS High Level Trigger 
Gianluca Comune
(Bern University)
On behalf of ATLAS Trigger/DAQ
High Level Trigger Group

2. Introduction
I will describe an online Event Selection Software (Steering) for the ATLAS High Level Trigger (HLT).
The Implementation draws on the current ATLAS offline Software.
The Steering is responsible for efficiently accepting or rejecting Physics Events.
The two key selection strategy choices taken are
Seeded Reconstruction approach.
Reconstruction in incremental Steps allowing early event rejection
These last two ingredients dramatically reduce the amount of CPU and bandwidth resources needed.

3. Overview Seeded reconstruction and Stepwise processing
means fast Event rejection
Seeded reconstruction:
Objects and Knowledge (Seeds) built by one algorithm can be used by another algorithm without the two being actually aware of each other and how the seeds were produced. This is achieved by use of Sequences and a data Navigation scheme
Processing by Steps:
The decision whether or not to reject an Event occurs in Steps. This decision is taken based on Signatures and on existing satisfied conditions signaled by the existence of certain Trigger Elements

4. The Steering Event Accepted Signature + STEP #2 + Algorithm Sequence
e50i+e50i ?
e50i
e50i +
isolation
isolation
STEP #2
Decision
e50
e50 +
Algorithm
Sequence
elecId
elecId
STEP #1
EM50
EM50 +
time
Trigger Element
Initial Seeds

5. The key Steering Components
Trigger Elements
History Objects
Algorithms
Data Navigation
Sequences
Signatures
Seeding and RoI
STEERING

6. Trigger Elements Trigger Elements provide access to Data Objects.
Steering creates and passes TEs to Algorithms.
“relation” TE
Data Object
Trigger Elements are linked to objects with a “relation”. Objects are external to the Trigger Element itself
The “relation” is encoded in objects called History Objects
Trigger Elements are used by algorithms to signal a satisfied Hypothesis to the Steering

7. History Objects
History Objects are the entities in charge of recording
the “relation” between TEs  Objects
History Object
key1
pointer<T1>
key2
pointer<T2>
….. ….
keyN
pointer<TN>
(T=any Object Type)
In our implementation History Objects are STL multimaps that record pairs keypointer<T>
The “key” can be any value. Its explicit meaning is not constrained by the History Class definition

8. History Objects: how they work
“seeded by”
0xBBBBBB TE TE
“uses”
“excludes”
0xCCCCCC TE TE
Type A
0xDDDDDD
Type B
History Object
“seeded by”
0xBBBBBB
“uses”
0xCCCCCC
“excludes”
0xDDDDDD

9. Algorithms Algorithms are responsible for the Reconstruction.
The Steering calls them on a conditional basis
The set of algorithms that can potentially be executed in a run is configured based on XML files (see M. Elsing’s talk)
Dynamical information in terms of existing Trigger Elements determine if and how many times a particular algorithm has to be executed
Algorithms signal success condition to the Steering by activating the input Trigger Element

10. Algorithms and Navigation
Receive
Input TE
Navigate
to Seed
Create New
Object
Link Object
to Input TE
Signal
Success
ALGORITHM
time
Navigate
“seeded by” TE TE
Input TE
“uses”
Link
“uses”
Object A
Object B
Seed
New Object

11. Sequences TE Algorithm TE
Sequences are “instructions” to the Steering on how
and when one or more algorithms should be executed TE
Algorithm TE
seed
hypothesis
Sequences are used by the Steering Sequencer to decide whether to execute an algorithm or not.
If the seed TE is active then the algorithm is executed with the hypothesis TE in input
Only Algorithms which are relevant for the Physics Signatures that we are looking for are executed

12. Signatures
Signatures are a combination of active TEs that identify
the signature of a desired physics process.
e50i + e50i
Signatures are used by the Steering Decision at every Trigger step to select wanted and promptly reject unwanted events
They are checked against the number of existing active TEs to determine if they are Satisfied. The ones that are fulfilled are recorded to be used later by the Steering Result
The Signature Table in a run is defined in configuration files and identify the Physics signals that will be searched

13. Seeding and RoI
The Steering Seeding mechanism is a natural implementation of the concept of Region of Interest (RoI) (see V. Boisvert talk).
The RoI is the perfect candidate as initial seed for the reconstruction chain.
An algorithm will be able to limit its processing only to this geometrical Region. Likewise the downstream algorithms can refer to the same Region through the Data Navigation
With the RoI mechanism, Algorithms request and process only a small fraction of the full Event

14. Active TEs & Signatures
Putting all together
Previous
Trigger Level
Steering
Initial Seed
Satisf. Sigs
Signatures
Sequences
Result
Decision
Sequencer
Yes/No
Exec (TE)
Config Sequences
and Signatures
Create Event Result
Retrieves Satisfied
Signatures
Active TEs & Signatures
Set up “seeded by”
Create input TE
Algs
Reco
Active
Config
Result TE
History A
“uses”
“seeded by” TE
Reco TE

15. Current prototype
The critical requirement is that the overhead introduced by the Steering framework should account only to few % of the 10ms reconstruction time available at Trigger Level 2.
Event
~1200μs
(5 signatures)
3 seeds
Write History entry
~16μs
O(10)/Evt
Read History entry
~38μs
Using RH7.3, optimized gcc-2.95 build, 1GHz/512MB machine. Times taken with NetLogger timing libraries
This is a first iteration and it looks like we are on the right track

16. Conclusion
We have a working Steering prototype based on the ATLAS offline Software Framework
We are in the process of integrating the Steering with algorithms and Data Access to exercise a full event selection chain for electron and photons
We expect the design/implementation choices to be reviewed mostly regarding the very tight timing constraints

17. Acknowledgments HLT Steering HLT Configuration Navigation
A. Radu (Bern Uni.), G. Comune (Bern Uni.)
HLT Configuration
T. Schörner-Sadenius (CERN), M. Elsing (CERN)
Navigation
S. George (RHUL), A. Lowe (RHUL), M. Grothe (CERN)