742
D. Froidevaux
now has access to the full event data. This farm runs the final, essentially offline-like
selection, “seeding” the reconstruction from the objects previously identified by the
L2 trigger in order to reduce the total processing time. The rate input into the event
filter farm is a few kHz, so the selection at this level has to provide typically a factor
of 10 in rate reduction.
The system relies on commercially available networks for the interconnection
between the readout buffers and the HLT farm. The advent of very inexpensive
Gbit Ethernet switching fabrics and processor interfaces, along with the rapidly
deployable 10 Gbit Ethernet standard, have rendered all early thoughts (back in the
mid-1990’s) of potential home-grown solutions obsolete.
16.6.2 L1 Trigger System
The L1 trigger has to process information from the detector at the full beam crossing
rate of 40 MHz. The very short time between two successive beam crossings (25 ns),
along with the wide geographical distribution of the electronic signals from the
detector, excludes real-time processing of the full detector data by general-purpose,
fully programmable processing elements.
The data are, instead, stored in pipelines awaiting the decision of the L1 trigger
within up to 3 μs. The maximum time available for processing in the L1 trigger
system is determined by the limited memory resources available in the frontend (FE) electronics which store the detector data during the L1 decision-making
process. Technology and financial considerations at the time of the design resulted
in a limit of at most 128 bunch crossings, i.e. the equivalent of approximately
3 μs of data, which can be stored in the FE memories. This total latency of 3 μs
therefore includes the unavoidable latency components associated with the transfer
of the detector information to the processing elements of the L1 trigger and with
the latency of the propagation of the L1 decision signals back to the FE electronics.
The resulting time available for the actual processing of the data is no more than
∼1−1.5 μs.
In order to avoid dead-time, the trigger electronics must also be pipelined since
every process in the trigger must be repeated every 25 ns. The high operational speed
and pipelined architecture also imply that only specific data can be brought to the
corresponding processing elements in the trigger system. In addition, the data must
flow synchronously across the trigger logic in a deterministic manner.
This architecture results in the presence of data from multiple crossings being
processed sequentially through the various stages of the trigger logic. To achieve
this, most trigger operations are either simple arithmetic operations or functions,
which use memory look-up tables, where an address is used to produce rapidly
a previously calculated (and stored) result. Moreover, the short time available
significantly restricts the data, which can be used in forming the L1 trigger decision,
in two ways: on the timing front, the only usable data can come from detectors with
very fast response or from slower detectors, which have both good time resolution
D. Froidevaux
now has access to the full event data. This farm runs the final, essentially offline-like
selection, “seeding” the reconstruction from the objects previously identified by the
L2 trigger in order to reduce the total processing time. The rate input into the event
filter farm is a few kHz, so the selection at this level has to provide typically a factor
of 10 in rate reduction.
The system relies on commercially available networks for the interconnection
between the readout buffers and the HLT farm. The advent of very inexpensive
Gbit Ethernet switching fabrics and processor interfaces, along with the rapidly
deployable 10 Gbit Ethernet standard, have rendered all early thoughts (back in the
mid-1990’s) of potential home-grown solutions obsolete.
16.6.2 L1 Trigger System
The L1 trigger has to process information from the detector at the full beam crossing
rate of 40 MHz. The very short time between two successive beam crossings (25 ns),
along with the wide geographical distribution of the electronic signals from the
detector, excludes real-time processing of the full detector data by general-purpose,
fully programmable processing elements.
The data are, instead, stored in pipelines awaiting the decision of the L1 trigger
within up to 3 μs. The maximum time available for processing in the L1 trigger
system is determined by the limited memory resources available in the frontend (FE) electronics which store the detector data during the L1 decision-making
process. Technology and financial considerations at the time of the design resulted
in a limit of at most 128 bunch crossings, i.e. the equivalent of approximately
3 μs of data, which can be stored in the FE memories. This total latency of 3 μs
therefore includes the unavoidable latency components associated with the transfer
of the detector information to the processing elements of the L1 trigger and with
the latency of the propagation of the L1 decision signals back to the FE electronics.
The resulting time available for the actual processing of the data is no more than
∼1−1.5 μs.
In order to avoid dead-time, the trigger electronics must also be pipelined since
every process in the trigger must be repeated every 25 ns. The high operational speed
and pipelined architecture also imply that only specific data can be brought to the
corresponding processing elements in the trigger system. In addition, the data must
flow synchronously across the trigger logic in a deterministic manner.
This architecture results in the presence of data from multiple crossings being
processed sequentially through the various stages of the trigger logic. To achieve
this, most trigger operations are either simple arithmetic operations or functions,
which use memory look-up tables, where an address is used to produce rapidly
a previously calculated (and stored) result. Moreover, the short time available
significantly restricts the data, which can be used in forming the L1 trigger decision,
in two ways: on the timing front, the only usable data can come from detectors with
very fast response or from slower detectors, which have both good time resolution