Present : F. Wickens, J. Schlereth, Y. Ermolin, S. Tapprogge, K. Korcyl, R. Scholte, P. le Du, N. Ellis, R. Blair, S. Wheeler, S. Gonzalez, P. Clarke, A. Waananen, K. Korcyl, M. Levine, S. George, J. Vermeulen

1. Updates of parameters, distributions for fragment sizes and execution times

S. George presented an overview of modelling parameters for the B-physics high-level trigger. For more details see the update of ATL-DAQ-2000-031 (to be released soon) (slides). The TRT scan takes a long time (400 ms on average on 600 MHz Athlon, with a tail up to more than 1.5 s), this is much longer than the time used for modelling (25 ms for 1000 MIPs processor). For scanning the Pixels and SCT track extrapolation on average 75 ms is measured on a 450 MHz Pentium III, which is roughly compatible with the 20 ms assumed for a 1000 MIPs processor. In the discussion following the presentation is was concluded that a smaller group (J. Baines, S. George, J. Vermeulen and others who want to contribute) should discuss and propose scenarios to be studied. The scenarios should include one with FPGA co-processors. The ATLAS week in October provides a natural point in time for a next modelling meeting in which further information can be discussed.

Next R. Scholte presented results on distributions for sizes of event fragments output by the SCT ROBs for different ROB mappings and for high luminosity running. Almost Gaussian distributions with a small tail for higher event sizes have been observed. The fragment size distributions for the SCT and the TRT will also be studied. It was noted that the data sample used may give a too pessimistic view. On the other hand within a RoI the detector occupancy may be higher than outside a RoI and that there may be correlations between ROBs. S. George will provide to R. Scholte a number of a tape with B-physics events, so that the same study can be repeated for low luminosity running.

Discussion :

- S. Gonzalez has sent an email to J. Bystricky and J. Vermeulen to draw attention to ATL-DAQ-2000-042 ("Further Studies and Optimisation of the Level 2 electron/photon FEX Algorithm"), discussing RoI sizes, ROB mappings, processing sequences etc. . P. le Du and S. Gonzalez will follow this up.

- An exact specification of the information provided by the LVL1 muon trigger is not available. N. Ellis suggested to send an email to L. Nisati.

- P. le Du noted that a possible compression of the calorimeter data should be looked at. He will discuss this issue with the LAr group.

J. Bystricky was not present, but P. le Du did present his slides later during the section on "computer modelling results". Issues mentioned included : for 40 and 75 kHz LVL1 rate different trigger menus should be used (not just scaled menus), the insatisfactory state of the knowledge concerning the muon RoI information generated by LVL1, the use of parameters derived from the testbed results, the use of secondary RoIs and the use of single or twin network architectures. J. Bystricky intends to work on further development of the paper model taking these issues into account. Two action points were identified : P. le Du said that J. Bystricky will look into how the trigger menus have to be updated (for which S. George is to be consulted) and into the question of secondary RoIs (in consultation with S. Gonzalez).

2. Computer modelling results

R. Blair compared the quality of ATM testbed performance estimates obtained with simdaq (using the parameters also used for the paper and full computer model studies) with testbed performance estimates by the Ptolemy group (using parameters tuned to the measurements) (slides). The quality of the latter estimates clearly is better.

S. Wheeler presented an overview of results from the Ptolemy group. Most of the material can be found in ATL-DAQ-2000-039. As already noted by R. Blair, good agreement has been found between the simulation results and the ATLAS testbed results. However, the results are for a system consisting of 1 ROB, 1 Supervisor and 1 farm processor (with 2 or 3 worker threads), a study of more complex systems is important for further testing of the validity of the parameterization. Work is also going on to model the full system. Results have been obtained for the first step of processing for the high luminosity menu. R. Cranfield recently joined the group and will take care of adding the remaining processing steps. S. Wheeler will leave the group. She has discussed with G. Lehmann integration of the EB model and the LVL2 Ptolemy model, which may be possible on a time scale of a few weeks. However, G. Lehmann will also leave, but her work will be continued by somebody else. P. le Du remarked that it is of interest to study the case where event building is done via the network which is also used for LVL2 traffic.

J. Vermeulen reported that G. Lehmann as well as H-P. Beck could not be present. However, he received the PhD thesis of G. Lehmann, containing a detailed description of her work, the day before the meeting and showed one figure from it with simulation and measurement results for 4 x 4 event building using ATM technology. Both types of results are in excellent agreement with each other.

Next J. Vermeulen showed a number of results obtained with simdaq for the full model and documented in ATL-DAQ-2000-035. The results demonstrate the desirability of minimization of queueing in the system. In particular optimization of the assignment of events to farm processors provides a possibility to achieve this. Results on averages and minima and maxima have been checked in detail against the paper models and good agreement has been found. The program (simdaq version 5) and all configuration and menu files used for ATL-DAQ-2000-035, as well as all histogram dump files with the results described in the DAQ note are available from the (pilot project) modelling page. J. Vermeulen also referred to a discussion he had on the day before the meeting with K. Korcyl concerning details of the Ethernet switch model and controlling the generic switch model available in simdaq in such a way so that it behaves in the same way as the Ethernet switch model. This seems to be possible, but needs to be tried.

3. Discussion on future work

A comparison of the Ptolemy and simdaq results was discussed. J. Vermeulen remarked that this can be done in two ways : (1) use the parameters and models agreed on for the full model TP studies and full sequential processing with Ptolemy, (2) use the testbed parameters, models and configuration with simdaq. It is intended to do both, good progress hopefully is possible before October. There was some discussion concerning whether the processing model used in Ptolemy (process threads) and the model used in simdaq (each step of the processing of a single event is done by an individual process, which receives and outputs its data via data buffers) would give rise to different results, in particular due to the limited number of active threads allowed.

It was remarked that updates on fragment sizes, etc. should be specified in a DAQ-note. Relevant information is available, an overview needs to be compiled, which is something to be done by R. Blair and J. Vermeulen in consultation with members of the PESA group.
 
 
 

Summary of action points with respect to input for modelling, as identified during the meeting :

1. J. Baines, S. George, J. Vermeulen and others who want to contribute discuss and propose scenarios for the B-physics trigger to be used for modelling,
2. P. le Du and S. Gonzalez look at the consequences for modelling of the results reported in ATL-DAQ-2000-042,
3. P. le Du discusses a possible compression of the calorimeter data with the LAr group,
4. J. Bystricky, in consultation with S. George, specifies 75 kHz LVL1 rate trigger menus and updates of the present 40 kHz LVL1 rate trigger menus,
5. J. Bystricky, in consultation with S. Gonzalez, proposes scenarios for using secondary RoIs,
6. An email is to be sent to L. Nisati requesting an improved specification of the information provided by the LVL1 muon trigger,
7. R. Blair and J. Vermeulen collect all information on fragment sizes, execution times, etc. for inclusion in a DAQ note presenting this information.

The time scale for these activities is set by the October ATLAS week.

Notes by J. Vermeulen, 21 June 2000