These notes currently only concern the presentations on modelling
Present : M. Abolins, A. Bogaerts, R. Blair, R. Bock, J. Bystricky, D. Calvet, R. Cranfield, R. Dobinson, M. Dobson, P. leDu, J. P Dufey, N. Ellis, Y. Ermolin, S. Falciano, S. Gonzalez, B. Green, M. Huet, R. Hughes-Jones, K. Korcyl, I. Mandjavidze, J.R. Hansen, C. Hinkelbein, R. Hughes-Jones, G. Lehmann, L. Mapelli, M. Mueller, B. Pope, F. Saka, R. Scholte, S. Tapprogge, J. Vermeulen, A. Wäänänen, P. Werner, S. Wheeler, F. Wickens
Notes by J. Vermeulen, 27 November 1999
Ethernet Switch Tests
R. Hughes-Jones, F. Saka
Gigabit Ethernet testbed
Ethernet switch modelling
K. Korcyl described a model of an Ethernet switch. Model results and measurement results are in good agreement. The model was used for studying the behaviour of a large switch built from smaller switches and fed with exponential random traffic from sources representing ROBINs to destinations representing LVL2 farm processors. The number of sources and the message sizes were taken from the paper model. The model and results have been documented : see the modelling page.
Ptolemy modelling of testbed
S. Wheeler described progress with modelling the Ethernet testbed and Ethernet switches in Ptolemy. A model of a cascadable switch with 4 ports is available, see the modelling page for details. While waiting for a more accurate model a "full system" representing the SCT barrel was built with the switch model available. This model functions but needs to be calibrated. With the help of the Ptolemy scripting language a larger system consisting of 1142 ROBs, 20 processors, and 581 cascaded 4-port switches was set up. One second of running with parameters obtained from the paper model can be simulated in 1 hr and 40 minutes wall clock time. See the modelling page for furgher information.
ATM testbed results
I. Mandjavidze (slides)
ATM testbed modelling
R. Blair used SIMDAQ with the standard parameters scaled with a factor 2.5 for reproducing the ATM testbed results. Qualitative results look OK, but the model needs to be calibrated.
Results of SCI measurements
F. Wickens
Modelling of the Event Builder
G. Lehmann reported on her work on modelling the Event Builder with Ptolemy. A new polling mechanism in the sources and destinations results in a better behaviour than previously observed. The credit assignment scheme in the DFM works OK. A 100 x 100 switch with 622 Mbit/s links should be capable of satisfying the requirements.
Paper model
J. Bystricky described his paper model study on grouping calorimeter ROBIns and work in progress on paper modelling of sequential processing of data from the different layers of the calorimeter.
R. Bock mentioned paper model results for the Active ROB Complex. These results are now documented in ATL-COM-DAQ-99-020.
J. Vermeulen discussed an update of the event fragment sizes and of paper model results, in particular with respect to the effect of using only LVL1 RoIs for requesting the data from the ROBs. The slides contain a short overview of relevant results, see for more extensive results the document available via the modelling page. After some discussion it was decided that it will be assumed that the TRT scan generates on average 2.0 RoIs for the calorimeter and the muon detectors. B-jets requiring analysis of jet RoIs in the tracker are not taken into account.
Full computer model
J. Vermeulen reminded the audience of some earlier results of the full computer model, and of the good agreement between computer and paper model results for average message and data rates and for processor occupancies.Some distributions which were not understood were shown (slides). One of the distributions showed that for the low luminosity trigger muon RoI data requested from the em calorimeter could need about 14 ms to travel from the ROB to one of the farm processors (explanation found after the meeting : the muon RoI data is requested at the same time as the inner tracker data for the scan and can become queued in the switch until all data from the tracker is input, which takes about 14 ms with the fragments sizes assumed for 15 MByte/s bandwidth per network link).
Discussion on the workplans up to Beatenberg and preparation for the workshop
Testbeds
Modelling
15 November : document by K. Korcyl on Ethernet modelling and update
of document on paper modelling results, see modelling
page for both documents,
29 November : other documentation updated.
Preliminary discussion on the workplans after Beatenberg
Testbeds
Modelling
In Beatenberg further updates of documents have to be discussed.
