740
D. Froidevaux
16.6.1 General Considerations
The most important architectural decision in the Trigger/DAQ system is the number
of physical entities, or trigger levels, which will be used to provide the rate
reduction of O(10 3 ) from the rate of 100 kHz accepted by the L1 trigger to
the final rate to storage of O(10 2 ) Hz. Current practice for large general-purpose
experiments operating at CERN, DESY, Fermilab, KEK and SLAC is to use at
least two more entities, colloquially referred to as the L2 and L3 triggers. Some
experiments even have a L4 trigger. The higher the level, the more general-purpose
the implementation, with the L3 and L4 trigger systems always relying on farms of
standard commercial processors.
The implementation of the L2 trigger system varies significantly across experiments, from customised in-house solutions to independent processor farms. The
issue encountered by all experiments, which have opted for multiple trigger levels,
is the definition of the functionality that the L2 system should provide. Of all the
trigger levels after L1, the L2 trigger is the most challenging one, since it has to
operate at the highest event rates, often without the benefit of full-granularity and
full-resolution data, though with data from more detectors and of higher quality
than that used by the L1 Trigger. Decisions that have to be made are the rejection
factor that the L2 trigger must provide, the quality of the information it will be
provided with, the interconnects between the detector readout, the L1 trigger and
the L2 trigger, and finally, the actual implementation of the processing units which
will execute the selection algorithms.
Ideally, the High-Level Trigger (HLT) should have no built-in architectural nor
design limitations other than the total bandwidth and CPU, which can be purchased
based on the experiments resources. Indeed, from very early on, the desire to
provide the maximum possible flexibility to the HLT led to the first design principle
adopted by ATLAS: the HLT selection should be provided by algorithms executed
on standard commercial processors, avoiding all questions and uncertainties related
to home-grown hardware processors.
The architecture is depicted schematically in Fig. 16.16. The implementation of
the L2 trigger has the advantage that much less data are required to flow into the
event filter farm, which in turn has more time to process incoming events. The L2
farm, on the other hand, has to provide a decision on all the events accepted by the
L1 trigger. To reduce the data flow into the L2 farm, only a fraction of the detector
information is actually transferred from the readout buffers to the L2 processors.
This is the concept of the “Region of Interest” (ROI). In brief, the result of the L1
trigger drives the L2 processing, by indicating the regions of the detector which are
involved in scrutinising the physics object (electron, muon, jet,. . . ) identified by the
L1 trigger. These regions are small, with a total data size of only a few percent of the
total event size, so that the full set of data from these regions can be transferred to
the L2 farm. The L2 algorithms employ sequential selection and usually not all the
data from the ROI in question have to be read in. This farm has tens of ms to provide
the L2 decision. The events accepted by L2 are sent to the event filter farm, which
D. Froidevaux
16.6.1 General Considerations
The most important architectural decision in the Trigger/DAQ system is the number
of physical entities, or trigger levels, which will be used to provide the rate
reduction of O(10 3 ) from the rate of 100 kHz accepted by the L1 trigger to
the final rate to storage of O(10 2 ) Hz. Current practice for large general-purpose
experiments operating at CERN, DESY, Fermilab, KEK and SLAC is to use at
least two more entities, colloquially referred to as the L2 and L3 triggers. Some
experiments even have a L4 trigger. The higher the level, the more general-purpose
the implementation, with the L3 and L4 trigger systems always relying on farms of
standard commercial processors.
The implementation of the L2 trigger system varies significantly across experiments, from customised in-house solutions to independent processor farms. The
issue encountered by all experiments, which have opted for multiple trigger levels,
is the definition of the functionality that the L2 system should provide. Of all the
trigger levels after L1, the L2 trigger is the most challenging one, since it has to
operate at the highest event rates, often without the benefit of full-granularity and
full-resolution data, though with data from more detectors and of higher quality
than that used by the L1 Trigger. Decisions that have to be made are the rejection
factor that the L2 trigger must provide, the quality of the information it will be
provided with, the interconnects between the detector readout, the L1 trigger and
the L2 trigger, and finally, the actual implementation of the processing units which
will execute the selection algorithms.
Ideally, the High-Level Trigger (HLT) should have no built-in architectural nor
design limitations other than the total bandwidth and CPU, which can be purchased
based on the experiments resources. Indeed, from very early on, the desire to
provide the maximum possible flexibility to the HLT led to the first design principle
adopted by ATLAS: the HLT selection should be provided by algorithms executed
on standard commercial processors, avoiding all questions and uncertainties related
to home-grown hardware processors.
The architecture is depicted schematically in Fig. 16.16. The implementation of
the L2 trigger has the advantage that much less data are required to flow into the
event filter farm, which in turn has more time to process incoming events. The L2
farm, on the other hand, has to provide a decision on all the events accepted by the
L1 trigger. To reduce the data flow into the L2 farm, only a fraction of the detector
information is actually transferred from the readout buffers to the L2 processors.
This is the concept of the “Region of Interest” (ROI). In brief, the result of the L1
trigger drives the L2 processing, by indicating the regions of the detector which are
involved in scrutinising the physics object (electron, muon, jet,. . . ) identified by the
L1 trigger. These regions are small, with a total data size of only a few percent of the
total event size, so that the full set of data from these regions can be transferred to
the L2 farm. The L2 algorithms employ sequential selection and usually not all the
data from the ROI in question have to be read in. This farm has tens of ms to provide
the L2 decision. The events accepted by L2 are sent to the event filter farm, which