Fix bug in new Moore documentation section
Compare changes
Files
2+ 8
− 2
@@ -3,7 +3,10 @@ Architecture of the LHCb Upgrade trigger
The detector data is received by 163 "event building" (EB) servers, each of which hosts up to three custom TELL40 readout boards. Each board receives data from one part of the detector. The EB servers are connected by a dedicated network which allows these event data fragments to be built into full LHCb events. This aggregation occurs inside the EB server memory, and both the amount of memory and the memory bandwidth must be sufficient to handle peak data rates. For this reason EB servers are equipped with 512 Gigabytes of RAM each. These events are then aggregated into "multi-event packets" (MEPs) before being passed on for further processing. The number of events or bunch crossings in a MEP is a tunable parameter of the system but is typically around 1000 events each. Passing MEPs instead of individual events minimizes I/O overheads in the system.
@@ -11,7 +14,10 @@ The first trigger stage (HLT1) is implemented in GPUs, which are hosted in the s
A subset of the events selected by HLT1 are themselves used as inputs to the alignment and calibration algorithms. To give just a few examples: a sample of unbiased collision events with hits in the VELO are used to align the VELO; events enriched in Cabibbo-favoured charm decays and :math:`J/\psi \to \mu^{+}\mu^{-}` decays are used to align the tracking system; a subset of events whose tracks are equally distributed across the RICH acceptance are used to align the RICH mirrors. The picture below shows a rough timeline of when each step of the alignment and calibration is executed relative to one LHC fill.
Once the detector alignment and calibration are available for a given run, the events selected by HLT1 are made available to the second trigger stage (HLT2) which is implemented on a farm of around 3700 CPU servers. At the same time these latest alignment and calibration constants are given to HLT1 and a new run is started. The disk buffer allows HLT2 processing to occur independently of whether the LHC is colliding or not, increasing the effective processing power of these servers. HLT2 outputs around 10 Gigabytes of data per second for offline processing. More details about the LHCb Upgrade dataflow can be found in the `Starterkit documentation`_.