SB's SVT analysis log

More in detail, count for each track how many roads have 1 combination to fit, 2 combinations, etc. It turns out for 4/4 there is no road with 5 or more, while for 4/5 is is much more corwded, also because roads are wider. Sometime will try to use Alberto's 128K pattern file with narrower roads to see if the problem is the road width more then the 4/5. Here are the plots of nroad, ncomb and nroad with 1-2-3-4 combinations, for the two runs:

If I only look at tracks that make no roads with 5 hits, so that the "usual proposal for ghostbusting before tf" can notbe applied, they are 30% of the event and the overall road/combination multiplicity there is a factor 2 less then the total 4/5 multiplicity, or 5x the 4/4. This is a likley irreducible load for the TF, that is about a factor 1.5 on the 4/4 timing.
Just for 4/5 run (164304 first 1K events). Top: number or roads and number of combinatio per XFT track, all of them. Bottom: only the cases in which there is no road with 5 hits for this track.

June 19 First try at a real ghostbusting before TF. Implement as a hack in svtsim_diagmodule the "perfect" ghostbusting and look at distributions. Assume that the following will be possible:

• roads will be classified according to what superstrips in the pattern are hit (no need to record hits, just superstrips)
• all 5-hit roads will be passed on to the TF
• all roads that have only 4 hit-superstrips will be suppressed if the same 4 superstrips are present in a 5-hit road
• in case more then one 4-hit roads share the same 4 hit-superstrips, only one will be passed, the others will be called ghosts and suppressed
• the TF result for 4-hit roads will only depend on the 4 hit-superstrips (and the wedge number) and not on the road number. I.e. the constants will have to be those for the z barrel of the first hit-superstrip, not those for the z barrel of SVX layer 0 in that road. This is not the case now.

1. Number of roads (ave=21)
2. Number of 5-hit road (ave=4)
   Note that in 30% of the cases there is no 5-hit road.
3. Number of 4-hit roads (ave=17)
4. Number of 4-hit roads whose 4 superstrips with hits are a subset of the the 5 superstrips of a 5-hit road (ave=10)
5. Same distribution as before, removing the bin at 0
6. Number of 4-hit roads whose 4 superstrips with hits are an exact match the the 4 superstrips with hits of another 4-hit road. (ave=3)
   Note that if there are 10 roads with the same 4 superstrip with hits, there is an entry at 9, this plot (like the previous 2) is the number of "busted" roads, one of the 4-hit roads has to survive to be fitted.
7. total number of ghostbusted roads (ave=13)
8. number of surviving roads to be fitted by TF (ave=8)

June 23 Run same code on Alberto's favorite run run152133 (run 152133. using his mapset/patterns mapset fnal:/cdf/ome/annovi/SVT/svtsim_new/svtsim/svtdata/offline_100030_20030613-my32k-1288812.mapset also saved in trieste:/home/belforte/svt/491/svtsim/svtdata
after having modified from a hack into svtsim_diagmodule.cc to the separate simulation of a new board (rgb) inserted on HB output. I obtain

Same plots, removing limit at 63 roads (Alberto has 126 in his mapset)

zooming in, still with road limit at 126

to be compard with Alberto's result:

so we have the same input (top plot=AMS output), but slightly different results after ghostbusting. Since results are very similar, we decide to leave the chase for the small bug to when it will matter.
and these are Alberto's plot with full range

July 9 Feasibility study for removing road duplicates before TF (road busting).

1. Overview and strategies on what to do in hardware. As a sketchy note
2. Study of what we can gain with a RB after the HitBuffer. There is probably still a small bug that Alberto discovered in my ascii ntuple making, where I do not count multiple XFT tracks in the same SS, so number of combinations is slightly underestimated. Should not affect results. I ran svtsim_diag modifying svtsim_diagmodule.cc to get my "ascii ntuple" and adding the RB simulation as/when needed (see point 3. below). I ran on run 164304, i.e. our first run with 4/5 but using the AMS ucode version 3.0 in svtsim, so to remove XFT tracks below 2 GeV as we do now in real life. In practice I used the mapset 4outof5_20030614141858 I made 4 plots:
   • Fit load in most crowded wedge "before the cure". From top left to bottom right:
     1. Number of Roads in the wedge with the most roads.
     2. Number of combinations to fit in the wedge with the most combinations to fit.
     3. Number of XFT tracks in the wedge with the most combinations to fit.
     4. Number of tracks in output from TF in the wedge with the largest TF output.
     5. Total number of tracks in output from all TF's (i.e. final GB input).
     6. Number of tracks in output from final GB (i.e. in input to Level2).
   • Fit load "after the cure". I.e. after inserting the RB between HB and TF. The comparison (looking just at histogram averages) suggest the folling gain:
     • From 30 to 12 roads
     • From 63 to 29 combinations to fit
     • From 23 to 7 tracks to final GB input
     • Modest loss of efficiency (3.5 -> 3.3 tracks final)
   • Details of road busting. From top left to bottom right: this time each plot is simply the combination of entries from all wedges with non-zero roads, i.e. not the most crowded one. In the end the information is the same.
     1. Number of roads in input.
     2. Number of roads with 5 hits.
     3. Number of roads with 4/5 hits.
     4. Number of roads with 4/5 hits that share the 4 hits with a 5-hit roads and can therefore be "busted"
     5. Number of roads with 4/5 hits that have the same 4 hits of another 4-hit road and can therefore be "busted"
     6. Number of busted roads (should be equal to sum of previous 2)
     7. Number of passed (OK) roads to Track Fitter
     8. Number of words read from HitBuffer (multiply by 40nsec to know much it takes to receive all the data from the HB)
     These plots more or less confirm previous conclusions: From 21 to 8 roads (30% reduction). Note that for this run with 4/5 patterns the average HB output load is 74 words = 3 microsec.
   • Details of road busting as before, but only look at wedges that have NO ROAD WITH 5-HITS. I.e. these are the wedges that would not be helped at all by a RB placed before the HitBuffer. These are about 25% of all wedges.
3. Simulation. I have modified svtsim to include an rgb board between HB and TF. All the files are in my local test release on pclx06.ts.infn.it. As of July 9 I tarred everything and copied it to FNAL in /cdf/spool/belforte/SVT/491/... I did all in 4.9.1 release.
4. Implementation. I described in Verilog the core functions of a RB placed after the HitBuffer, and fitted it using Quartus on the GB Apex. It looks great (47% usage, 68MHz clock). See details.