Present : R. Blair, R. Bock, J. Bystricky, J. Dawson, N. Ellis, Y. Ermolin, K. Korcyl, J.R. Hansen, R. Hughes-Jones, G. Lehmann, M. Nomachi, B. Pope, R. Scholte, J. Strong, S. Tapprogge, J. Vermeulen, F. Wickens

Via telephone : P. Clarke

Notes by J. Vermeulen

1. Parameters update (J. Vermeulen) (slides)

The  presentation of the last meeting was updated. In particular new numbers for the average sizes of the event fragments stored in the ROBs for the pixels and SCT from Nick Ellis led to a considerable reduction in the amount of data to be transported for the low luminosity trigger. In the discussion P.Clarke mentioned that the new version of DAQ note 62 is finished (this version has been made available after the meeting as ATLAS-COM-DAQ communication. On the basis of this new version some event fragment sizes had to be changed again, most notably for the SCT for the high luminosity trigger. The slides reflect the status at the time of the meeting).

2. Paper model

J. Bystricky mentioned the intention to model the electromagnetic presampler and the front, middle, and back layers of the electromagnetic calorimeter as separate devices. This makes it possible to request data from these "subdetectors" individually as would be needed for  a an algorithm processing the em calorimeter date in several steps, while reducing the rate after each each step. The complete model was not yet finished at the time of the meeting. He also presented results (slides) on a calculation of the average number of RSIs hit per RoI and on the average request rate per RSI for different grouping factors. Dependances were shown of the mean numbers of RSIs hit by 0.4x0.4 and 0.8x0.8 RoIs on the numbers of ROBs in each RSI for the em and hadron calorimeters. The functions may easily be fitted by two exponentials. The occupancy (percentage of RoIs hitting an RSI on average) is a fairly linear function of the number of ROBs per RSI.J. Vermeulen noted that the calculations were similar to the calculations done with his ROBsPerRoI program.

J. Vermeulen discussed (slides, same set as for 1.) the main consequences of the parameter updates for the paper model results and also showed how the number of processors required for the low luminosity trigger depends on the execution time of the B-physics trigger algorithm. This is of interest in view of the possibility to use special-purpose PCI-based co-processors for compute-intensive parts of the algorithm.

NB : Before completion of these notes it has become clear that the two different calculations of the average number of ROBs or RSIs per RoI are based on different interpretations of the information provided by the LVL1 trigger for the e/gamma/hadron and for the jet RoIs. From information provided by A. Watson it is concluded that the LVL1 trigger provides the location of the sliding window (dimensions 0.4 x 0.4 in eta-phi space for the e/gamma/hadron RoIs and 0.8 x 0.8 in eta-phi space for the jet RoIs)  and not the location of the local maximum within that window. This affects the paper and computer model results of J. Vermeulen, but in a favourable sense : RoI request rates are reduced.

3. Testbed models

R. Blair (slides) reported on the status of work on modelling an ATM test set up consisting of 8 nodes and 2 supervisors, interconnected by means of  ATM and a FORE switch. This was not yet succesful. It was found that the newest version of SIMDAQ (pre-release of version 4.5) crashed for the configuration to be studied.

NB After the meeting it was found that the code in SIMDAQ required for handling this configuration was incomplete, which also was responsible for the crash. The code has been updated in the mean time.

K. Korcyl reported (slides) on his work on a parametrized model of an Ethernet switch with OPNET 6.0. He found that "filler frames" may cause a drop in the throughput of the switch. It is the intention to reduce first the number of OPNET kernel calls in the simulation program and then to move to Ptolemy. Click here for more information. Probability density functions describing the behaviour of the switch are available

P. Clarke reported on the status of the development of the Ptolemy model of the Ethernet testbed. S. Wheeler and G. Crone have been working on the components around the switch, R. Hughes-Jones looks after the model of the Ethernet switch and maintains contacts with K. Korcyl. Configurations to be studied and measurements to be made are to be defined.

R. Bock remarked that the scattering of information in the ROBs forms a problem with off the shelf hardware and software.With a large enough ROB complex feature extraction could be feasible, but in any case RoI fragment building would imply much reduced message frequencies in the LVL2 network. Click here for more information. The intention is to set up a 4 processor system with ROBIns connecting via PCI. The effect of grouping is to be studied with paper and computer modelling with help from S. Wheeler and J. Vermeulen.

4. SIMDAQ update (J. Vermeulen) (slides, same set as for 1.)

With the current version of SIMDAQ (pre-release of version 4.5) it is now also possible to model event building for individual ROBIns connecting to the Event Builder switch. Also support for ports and switches working with packets (in the form of simple ATM ports and switches) has been added. The program had been run with the standard pilot project configuration (with a LVL2 farm of 768 processors) and set of parameters and with the new reduced values for the event fragment sizes for the SCT and pixels. Stable operation at the nominal LVL1 rates (40 kHz) has been observed for the low as well as the high luminosity trigger. Distributions for the LVL2 decision times and event building times (via 15 MByte/s links into 512 destinations and for a LVL2 accept fraction of 5 %) were shown.

5. Next Meeting and workplan

The next meeting was suggested to be held at the end of October or in the beginning of November, probably close in place and time to the next ROBComplex meeting.

NB In the Pilot Project Coordinator meeting of September 15 the decision was taken to have a combined testbed and modelling meeting in the morning of Thursday 4 November and separate testbed and modelling meetings in the afternoon of the same day in CERN.

It was planned by the convenors to have a look at the workplan after the integration meeting and the meetings in the ATLAS week following this meeting.

NB It was decided by the convenors to suggest two dates before the next meeting for receiving documents with a description and results of the work done. The first date would be 6 October for first versions or outlines (see the modelling page), the second date is 27 October for more complete documents, in time for people to digest before the next meeting.