Some hints to speed up the PAW Ntuples analysis

When the binning of the resulting histogram is known, use the parameter IDH in the command NT/PLOT. This will avoid one loop over the ntuple to compute the minimum and maximum value of the plotted expression. Example:

Instead of doing:

         PAW > NT/PLOT 1.X

Do:

         PAW > 1DHISTO 10 ....
         PAW > NT/PLOT 1.X IDH=10

Use the VMEM command to increase the size of the memory cache.
When, during a ntuple analysis, a lot of variables are not used, column wise ntuples are more efficient. With these kind of ntuples only the used variables are loaded from disk. This is especially true with "collaboration ntuples" which, very often, contain hundreds of variables.
When a selection mechanism uses fortran code (via the COMIS interpreter), use compiled code. Example:

Instead of doing:

        PAW > NT/PLOT 1.X select.f

Do:

        PAW > NT/PLOT 1.X select.f77

In the second case, the local fortran compiler is invoked. Note that in both cases the selection function is in the file ``select.f''.

This is a real example (with PAW version 2.07 and after) just to figure out what the speedup could be. Data are stored in a ~100Mbyte Column Wise Ntuple. The first command loops over the 378340 events and selects them according to an interpreted FORTRAN function. When the IDH parameter is used the elapsed time is reduced by a factor 2. If the FORTRAN code is compiled, there's a further speedup by 5, so the total gain factor is ~12 (from 291 to 25 seconds).

        PAW > nt/pl 10.px4 invm.f(12.)
         Real time 00:04:51, CP time 283.710


        PAW > nt/pl 10.px4 invm.f(12.) idh=22
         Real time 00:02:20, CP time 140.190


        PAW > nt/pl 10.px4 invm.f77(12.) idh=22
         Real time 00:00:25, CP time 22.660