Timing studies were performed on a T800 transputer accessed through a CAPLIN QT0 interface from a micro-VAX. We used a software harness [16] which allowed the overall program framework to run on the VAX, while the segment finding algorithm - coded in FORTRAN - was run on the transputer. The times were accurate to within the clock cycle of 0.064 ms. Only masks containing greater than 3 and less than 26 hits were considered for segment finding and no fiducial cuts were applied. A hit mask was defined to be one that contained more then three hits, while a hit crate was defined to be one that contained at least one hit mask. Table 8 shows the mean times taken by the segment finding code for the different event types.
Table 8: Mean time taken to find segments.
The distributions of the times taken to find segments in masks for the different event types are similar. Each distribution has a range extending from approximately 0.3 ms to 3.0 ms per hit mask, and a peak at 0.8 ms with a small shoulder just above the peak. Figure 10 shows the times required to find segments in hit masks for BG events.
Figure 10: Times taken to find segments in
masks.
Segment finding is performed in each crate separately by working on each mask in the crate one after the other. Hence, the interesting timing is the time to find the segments in all the masks in a crate. The distributions of the time per crate for the different event types are also similar, and figure 11 shows the times required to find the segments in all crates for BG events. A long tail is observed for each event type with the most probably time being about 0.8 ms per crate (which corresponds to approximately one mask per crate being hit). The large number of entries at zero in figure 10 show that some crates, and hence sectors, will often have no hits in them. The undesirable long tail in figure 11 may be due to events with, either most masks in a crate being hit, or a few masks in a crate having a large number of hits (or both). Studies to investigate the tail of the distribution are in progress.
Figure 11: Times taken to find segments in all
crates.
The mean of the distribution in figure 11 (last row of table 8) can not be much longer then 1 ms or the deadtime in the SLT will be large. At the moment, the code is approximately a factor of four too slow, but an increase in speed of about a factor of four is expected when the algorithm is coded in occam and further optimization of the code is performed. occam coding and optimization are in progress.