Michigan: M.Abolins
Argonne: R.Blair
UCL: R.Cranfield
Cracow: K.Korcyl
NIKHEF : J.Vermeulen
CERN: S.Wheeler, P.Golonka
1. Paper modelling: The C++ program replacing the spreadsheet has been made available via the modelling web page. M.Abolins will run the program for the set of parameters provided by V.Vercesi.
2. ROS modelling: R. Cranfield and J. Vermeulen exchanged emails on the model to be used for the performance testbed. The conclusion is that the existing models with appropriate parameter values probably can be used. However, at2sim is lacking the "ROBOuts" of simdaq, i.e. components that are intermediary between small groups of ROBIns and the network, so these may have to be added to at2sim.
3. Data Collection modelling and HLT modelling:
S. Wheeler: the calibration measurements for the DFM component need to be redone, followed by an update of the DFM model parameters. The SFI testplan is not yet finished, this has to wait till August for completion. It is still not clear who will take care of the calibration measurements.
K. Korcyl finished a first draft of a document with a description of the parameterized switch model. He also described the results of a study together with Anna Kaczmarska aimed at reducing the execution time of the strawman system model. The main lesson is that the number of events (simulation events, occurring in the discrete event model) in the event queue should be minimized. A "star" should not book more than one event per time and refire to itself if more than a single event is needed. J. Vermeulen remarked that in simdaq the same approach is used and also that further optimization is possible by a suitable choice of the implementation of the event queue and of the event queue handling (see simdaq documentation). K. Korcyl replied that he does not want to modify the internals of Ptolemy. The at2sim program was modified taking into account the findings. For the strawman architecture with the full ROB system, 250 LVL2 processors, 150 SFIs and 112 switches (simplified model) running at a LVL1 rate of 32 kHz and for sequential processing in LVL2, 4 hrs of computing time are needed to simulate 1 second of running time on a 1 GHz Pentium III machine with 1 GByte of memory. With the parameterized switch models 6 hrs are needed to simulate 1 second of running time. The optimization problem described in the previous meeting still exists (i.e. compilation with optimization switched causes at2sim to produce incorrect results). It is estimated that a reduction of about 25% in running speed could be obtained if optimization could be used. Finally K. Korcyl mentioned that the strawman document was updated. The new version was sent to R. Dobinson and M. Levine, replies had not yet been received at the time of the meeting.
P. Golonka reported that the new model of the L2PU basically works. He started writing the the documentation, a draft should become available afther the July Trigger/DAQ week. The large scale tests on the LXCERN cluster were also mentioned by P. Golonka, for performance modelling these are not of interest as no timing measurements have been done.
4. at2sim: No further news than reported under 3.
5. Simdaq: Simdaq has been run for the "standard" paper model parameters for low, intermediate and high luminosity and the "least queued" LVl2 processor assignment strategy. After removing some errors in the configuration and menu files good agreement of average values with those calculated with the paper model (spreadsheet or "paper" program) has been observed. A spreadsheet with decision time distributions (for about 80 % LVL2 processor utilization) and output of simdaq concerning averages is available.
6. Chiba City: A first step in the direction of running the new Data Collection software on the system has been taken by installing the software on R.Blairs laptop. Reiner has set up scripts for the LXCERN cluster tests that from 7 lines of text generate about 7,500 lines. This could be of interest also for tests with the Chiba City system.
7. Workplan: In the new management structure the work on modelling is to be represented by J. Vermeulen in the TDSG and in the DAQSG. It seems desirable to focus the attention now primarily on modelling of the performance testbed and of an extrapolated full system, as foreseen in the original workplan Work is well under way and the proposal is to continue in the same way as before, if the steering groups do not disagree with this proposal. Monthly telephone meetings, organized in the same way as before, to discuss progress and results, are foreseen .
8. Next meeting: The next telephone meeting is planned for Tuesday 3 September, 15.30 h CERN time.
J. Vermeulen, 17 July 2002