Notes modelling working meeting February 27, 1997, CERN, and agreed action points Jos Vermeulen, March 1, 1997 The discussion concentrated on various numbers that are relevant for paper modelling, modelling (i.e. system simulation) and emulation. These numbers will be available from a document edited by Simon George. The URL of a draft version of it is : http://www.rhbnc.hep.uk/. The following items have been discussed : - Trigger menus, rates The RoI's corresponding to items in the minimal menu's are "primary RoI's", all additional RoI's in items in the extended menu are "secondary RoI's" The exclusive rates in the menu's are our working assumption, until more reliable values are available. In the high luminosity minimal menu the MU6 + EM15I item becomes : MU6 + EM20I. The exclusive rate for this item is 3000 Hz. In the high luminosity extended menu the J150 item has an exclusive rate of 0 Hz with the LVL2 missing energy trigger enabled, and an exclusive rate of 150 Hz with the LVL2 missing energy trigger disabled (note : this is different from what was said during the meeting, but is the consequence of the trigger menu's as provided in draft note <97/02/11 19:03:25 george>) In the low luminosity extended menu the J100 + J100 item has an exclusive rate of 1500 Hz with the LVL2 missing energy trigger enabled, and an exclusive rate of 1800 Hz with the LVL2 missing energy trigger disabled In the EM20I + EM20I + ... total exclusive rate there is a difference between minimal and extended high luminosity menu. (4000 vs. 3930 Hz). The same holds for for the MU6 + MU6 + ... total exclusive rate (970 vs. 1000 Hz) and for the low luminosity menu for the EM20I + ... total exclusive rate (9600 vs. 10000 Hz), and for the TAU80 + ... total exclusive rate (4960 vs. 5000 Hz). Simon will try to resolve this (ACTION POINT 1). (Not discussed in the meeting : in the draft note <97/02/11 19:03:25 george> the total rate for the extended high-luminosity trigger menu does not take the MU6 + EM20I item into account). - Trigger menu's, order of sequential processing and associated reduction factors It became evident from the discussion that the order of processing and the reduction factors need further clarification (ACTION POINT 2). - Event generation on the basis of the trigger menu's Assuming a certain distribution of the RoI's in eta (can be taken to be flat in first approximation) and no correlation between the different RoI's for a single event, it is possible to generate (many) events. A procedure for achieving this has been implemented in the "mapper" program by Reinier Dankers. It was agreed that events produced in this way can be used for obtaining first results in modelling and emulation studies and with the demonstrators. Events produced by ATRIG can then later be used for obtaining more reliable results. - Areas in eta-phi space "seen" by the ROB's For the forward and backward muon precision and trigger chambers there is no information available (ACTION POINT 3). For the barrel trigger chambers the mapping to ROB's is also not clear, although it seems most likely that the full length of the barrel and about 0.4 in phi is mapped onto a single ROB (ACTION POINT 4) - Possible centre positions for muon RoI's The offset of the centre of the grid associated with the possible centre positions of the muon RoI's with respect to eta = 0, phi = 0 is not clear (ACTION POINT 5). - Jet Trigger The jet trigger covers an eta range from -3.2 to 3.2. The centre position of the jet RoI steps from -2.6 to 2.6 with steps of 0.4. These numbers are different from the numbers in the table on page 19 of DAQ-note 62. The paper models therefore now also need to take into account the 16 ROB's on which the two "inner wheels" (terminology of DAQ note 62) (eta range -3.2 to -2.5 and 2.5 to 3.2) of the e.m. calorimeter are mapped. For the hadron calorimeter no additional ROB's have to be taken into account (ACTION POINT 6). - RoI sizes The sizes for muon RoI's are larger in the TRT and in the calorimeter than in the muon detector to compensate for the effect of the magnetic field. For muon RoI's the average number of ROB's per RoI therefore has to be recalculated. (ACTION POINT 6). - Number of ROB's hit per RoI For the calorimeters (when e.m. and hadron calorimeter are considered as separate subdetectors) the number of ROB's hit per RoI does depend only slightly on the position of the RoI. However, the number of ROB's hit per RoI is considerably higher for the TRT endcaps than for the TRT barrel. For this reason these are considered in the NIKHEF spreadsheet as two different subdetectors. The RoI rate per subdetector is taken to be equal to the total rate times the fraction of eta-phi space covered by the part of the TRT considered. - Data volumes RoI fragments output by ROB's For the TRT and for the SCT it was agreed that average numbers can be used, as fluctuations (due to zero supression) will occur for each ROB. However, for the calorimeter the data volumes of the RoI fragments are dependent on the ROB's. It was agreed that paper model calculations should be done with an average value and a maximum value. (ACTION POINT 7). The data volumes for the TRT and for the SCT need to be checked (ACTION POINT 8). - Processing times The TRT and SCT FEX execution times need to be updated (ACTION POINT 9). For modelling and emulation also the distribution of execution times is important. A suitable distribution should be asymmetric and have a long tail extending to long execution times. A choice of a distribution has to be made (ACTION POINT 10). - Process models At the end of the meeting also a discussion on process models was started, in view of CPU resources to be attached to the different process steps executed by the various processors of the LVL2 system. It became clear that a further discussion of the process models is necessary (ACTION POINT 11). The discussion continued the day after the meeting between Simon George, Nina Madsen and JV, and it is believed that the outstanding questions have been resolved. The results are provided below in textual form and are also available in the form of "bubble diagrams" Overview of action points ========================= 1. Investigate inconsistent rates in minimal and etended trigger menus (Simon George), 2. Determine order of sequential processing and associated reduction factors (No names mentioned, but suggest :Simon George, Dick Hubbard), 3. Fix ROB mapping for forward and backward precision and trigger muon chambers 4. Fix ROB mapping for barrel trigger muon chambers, 5. Determine offset of the centre of the grid associated with the possible centre positions of the muon RoI's, 6. Incorporate "inner wheels" of e.m. calorimeter in paper models for jet triggers. (Simon George, Marc Dobson, JV), 7. Use an average value and a maximum value for the calorimeter RoI fragment size in the paper models (Simon George, Marc Dobson, JV), 8. Check the data volumes for the TRT and for the SCT (Simon George and Marc Dobson), 9. Update the TRT and SCT FEX execution times (Simon George and Marc Dobson), 10. Choose a probability distribution for the execution times (Rudy Bock and JV), 11. Fix the processing models (Simon George, Nina Madsen, JV) and implement the correct models. Present view of the processing models ===================================== All times mentioned are CPU utilization times. For input and output a standard value of 50 microseconds is used. In the paper models this value can be set to a different value for the whole model and is referred to as "alpha * T0". For simplicity below an explicit value (the standard value) is used. ROB - Per decision block of 100 decisions 100 microseconds of CPU time is needed for management of the input buffer. Decision blocks arrive asynchronously and need to be buffered. For receiving a block and storing it 50 microseconds per block is needed - Per RoIR 10 microseconds of CPU time is needed for searching either for the data or for the RoIR, depending on which arrives first and for associated error checking. For receiving a RoI and storing it 50 microseconds per RoIR is needed. - For extracting the LVL2 data (probably a pointer to the LVL2 data) 10 microseconds per RoIR is needed. This 10 microseconds is to be added to the time needed for pre-processing. - For output of a RoI fragment 50 microseconds per fragment is needed - Output of event fragments to the LVL3 system is not considered for the paper models - Input of data from the optical fibre is done autonomously by dedicated hardware without processor intervention. - The input and output processes run asychronously and cause interrupt handling, the rest of the software is assumed to consist of a continuously repeated loop, i.e there are no context switches other than the invocation of interrupt handlers RSI The RSI has two functions : (A) receiving fragments from ROB's and merging fragments from one RoI into one larger fragment, and (B) : receipt of RoIDataRequest and distribution of these requests to the ROB's connected to the RSI. For each fragment input 50 microseconds is needed. It is assumed (warning : this has not been discussed !) that each ROB connected to the RSI receives its RoIR messages via the RSI. In this way the RSI will know for each event how many fragments can be expected from the ROB's. The merging of fragments into a single fragment proceeds with a speed of 50 MByte / s. For each message output 50 microseconds is needed. For each RoIR output and for each RoIDataRequest input also 50 microseconds is needed. SFI/FEX The function of the SFI is to receive all event fragments that are needed for a single feature extraction step and merge them into a single fragment that can be input to the FEX algorithm. This function can also be performed by the FEX itself. Per fragment input 50 microseconds is needed, the merging of the fragments proceeds with 50 MBytes / s. The SFI or FEX also receives in a pull model from the supervisor requests for feature extaction and distributes these requests to the RSI's. For each message input and for each message output again 50 microseconds is needed. Global The global processing step follows the model as described in the draft document from Simon George. Supervisor The supervisor follows the model as described in the draft document from Simon George.