Strategy of event building

1
Strategy of event building

• Sergey Linev, GSI

2
General aspects

• About 1000 outgoing from FEE data streams, each
  about 1 GBytes/sec
• Inside data stream each portion of data (single
  hit or hits cluster) has timestamp
• Each data stream is stored in active buffers for
  relatively long period of time (100 ms ? 1 s)
• About 1000 event builders, connected to active
  buffers over 1000x1000 switch network
• Data rate over switch is about 1TByte/sec

3
Event tagging and building

• Build multiplicity histogram over time for one or
  several layers of STS
• Position of peaks on multiplicity histogram
  should correspond to events times
• Event tagging assigning event time with
  respective id (tag number)
• Event building collecting data, correspondent
  to the same event in one computing node to
  perform event filtering

4
Idealized event building scheme
1 TByte/sec
Divide data on time frames and assign them to
event builders
1.1 TBytes/sec Barrel shift to balance switch
Switching network
Combine frames, produce event tags, build events
Provide events to L1 triggers
5
Idealized event building scheme

• All 1000 data streams have constant data rates
• All event builders always functional and capable
  to process all events in time
• Each buffer divides its data by constant time
  frames
• Simple round robin (Barrel shift) algorithm can
  be used to distribute data over switch network
• Event tagging is done after time frames from
  different streams collected in one event builder
• No need of central arbiter (controller)

6
Problems

• Variation of data size inside time frame because
  of
• Variation in interaction intensity (beam current)
  
• Variation of interaction type (central/peripheral
  collision)
• Detectors noise
• Accessibility of event builders
• Different kind of events required various time to
  be processed
• Failure or dysfunction of event builders
• These kind of problems can be treated by central
  controller

7
Time frames based switching
1 TByte/sec
Send data sizes in time frames to controller
Send time frames to builders according schedule
Calculate distribute schedule over all buffers
Switching network
Combine frames, produce event tags, build events
Send builders status (load) to controller
Provide events to L1 triggers
8
Time frames based switching

• Reasonable time frame size 10?50 µs
• Overlap between time frames should be equal to
  time definition of worse detector (for MAPS -
  5µs). Bigger overlap gt bigger additional data
  rate over switch
• All original data including detectors noise plus
  time overlap should be transferred over switch
• Scheduler can limit data transfer over switch
  only by excluding some time frames
• But another possibility is to enforce event
  tagging already on active buffers level

9
Event based switching
1 TByte/sec
Calculate local histograms and transfer them to
histogramming unit
About 10 of bandwidth required for histogramming
data
Combine histogram from 10 buffers, find peaks and
send them to controller
Tag events, distribute tags, calculate
distribute schedule over all buffers
Send data for event groups to builders according
schedule
Switching network
Send builders status (load) to controller
Combine frames, produce event tags, build events
Provide events to L1 triggers
10
Event based switching

• Histogramming will produce additional 70?100
  Mbytes/s traffic (7?10 to data rate) per channel
  
• About 3?5 histogramming unit required to produce
  histogram over one-two STS layers
• Events should be combined to groups of 100?500
  events
• About 50-80 detector noise can be filtered out
  by early event tagging mechanism
• Simple event filtering, based on STS
  multiplicity, can be applied

11
Scheduling algorithm

• Required to balance main switch network to avoid
  conjunction and lost in performance
• Simple round-robin (Barrel shift) in case, when
  size variation is negligible (less then 10)
• More sophisticated sorting algorithm, when size
  variations bigger then 10
• Band-width occupancy of 1000x1000 switch
  network

Size variations (2s) 10 20 50
Barrel shift 86 75 55
Sorting 96 95 93
12
Conclusion

• MAPS detectors will force to use frame based
  switching
• Event based switching allows to reduce detector
  noise and perform based on multiplicity event
  filtering before general switch network
• Histogramming produces additional traffic, which
  probably can be transferred over general switch
• Central scheduler is required to balance traffic
  and event builders loading
• To avoid lost of switch network bandwidth
  complicated scheduling algorithm should be applied