Throttling – Revisited Walter F.J. Müller, GSI, Darmstadt CBM Collaboration Meeting 13 th April 2010
13 April 2010 CBM Collaboration Meeting -- Walter F.J. Müller, GSI 2 Shown on CBM FEE/DAQ Workshop on 23 rd Feb 2010 Behavior under Overload Question: What happens if more data is generated in the FEE ASCIs than can be transported to and processed by the DAQ system ? part of the data all the time Default: Each FEE/ROC drops data independently Under overload we get part of the data all the time all of the data part the time Goal: All FEE/ROC drop data in a coordinated way Under overload we get all of the data part the time
13 April 2010 CBM Collaboration Meeting -- Walter F.J. Müller, GSI 3 Proposals Received Global approaches 'throttle-OR' wfjm Local approaches 1. 'Use drop slots' R. Szczygiel 2. 'Discard epochs in a prioritized way' S. Linev 3. 'Delete overhang data at begin of epoch' P. Fischer 4. 'fixed frame size' L. Schmidt First summary local methods clearly preferential to global approaches method 1+2 handle whole epochs need a substantial buffer space to work effectively perfect for later stages method 3+4 aim to concentrate data losses towards the end of a each epoch can be implemented on small buffer footprint
13 April 2010 CBM Collaboration Meeting -- Walter F.J. Müller, GSI 4 Canonical Architecture Data not time sorted PreAmp digitization FIFO Readout control, e,g, Token ring Data time sorted
13 April 2010 CBM Collaboration Meeting -- Walter F.J. Müller, GSI 5 Local Throttle Algorithm I PreAmp digitization FIFO Readout control, e,g, Token ring Add a THROTTLE Flip-Flop - set when FIFO full - cleared at epoch Data dropped if FF set Unclear what the system dynamics is !!
13 April 2010 CBM Collaboration Meeting -- Walter F.J. Müller, GSI 6 Local Throttle Algorithm II PreAmp digitization FIFO Readout control, e,g, Token ring Add a hit counter - data dropped if counter > limit - clear counter at epoch This implements 'fixed maximal bandwidth' per channel. Simple, clean, predictable, but potentially not optimal bandwidth usage.
13 April 2010 CBM Collaboration Meeting -- Walter F.J. Müller, GSI 7 Summary This is meant to start the discussion no pre-mature conclusions at this point This is meant to start the discussion no pre-mature conclusions at this point Discrete event simulations urgently needed to understand the system dynamics of different approaches Goal: a system with a benign behavior when operated close to the bandwidth limit despite some beam intensity fluctuations gradual increase of 'effective dead time' data stays normalizable Alternative: Substantial 'safety factor' in all bandwidth estimates...
13 April 2010 CBM Collaboration Meeting -- Walter F.J. Müller, GSI 8 The End Thanks for your attention