Trigger modifications to allow connections to remote farms

The worldwide trigger is closely related to event streaming, even though they fundamentally are different concepts:

• Streaming is intended to categorize events according to their physics type, which allows the experiment to prioritize their offline analysis efforts. Various streaming models have been discussed in the ATLAS "Streaming Working Group" and opinions on the topic widely vary. The working group proposed a baseline streaming model in the Trigger&Physics week of May 2006.
• The worlwide trigger collects geographically distinct resources to perform the last stage of event selection. This introduces a natural hierarchy in the Event Filter. The most reliable resources are located close to the experiment, but only available in limited numbers. They should be assigned to trigger only the highest priority data. Remote farms can be much larger, but will be fault prone. In the first stage of the experiment these farms should be assigned to trigger calibration channels only. Later, their task might be extended to trigger physics channels with a complicated signature, allowing a larger CPU budget for these classes of events.

A number of improvements in the Event Filter facilitate the exploitation of remote farms in the Higher Level Trigger.

• Within the event filter, calibration events will be processed by a different processing task then the physics events. A routing task assigns the events to the appropriate processing task based on a RoutingTag. More details can be found in an update of ATLAS note ATL-DQ-ES-0076 "Routing of events in the HLT". The same mechanism can be used to route events out of the event filter to remote farms.
• The communication protocols between SFI, EFD and SFO have been updated to allow for parallel connections. These parallel connections will be required when the average transport time of an event becomes small in comparison to the latency. This can either occur with the transport of partially build events on a local area network, or with fully build events on a wide area network. Details of the communication protocols can be found in ATLAS note ATL-DQ-ES-0040 "EFIO: Protocol specification".
• We are now modifying the InputTask of the EFD, responsible for fetching of the events, to enable parallel connections to the SFI. A new task, InputTaskMT, will provide this extension.
• The occupancy of the SharedHeap used to be controlled with a switch based backpressure mechanism. When the SharedHeap occupancy reached a predefined threshold, it simply suspended the request of new events. Although this mechanism works fine from EFD point of view, it introduces oscillations in the output buffers of the second level trigger. This instability mainly manifested itself with small events and large latencies. The backpressure mechanism of the SharedHeap has now been changed to a dynamic system, where the delay between the event requests is tuned to obtain a stable SharedHeap occupancy.

Contents

• Home
• News
• Planning
• People
• Interested?
• Subprojects
• Calibration
• Analysis
• Trigger
• Remote
• Grid
• Twiki
• Papers and
  presentations

External

• NIKHEF
• CERN
• ATLAS
• TDAQ
• Streaming
• PDP

A worldwide distributed trigger system