|
|
Quick tests comparing the performances of the PAW Ntuple Query Processor since version 2.07 with previous versions, show an average performance decrease by a factor between 2 and 3, depending on the platform, the job etc.
More detailed tests show that reality is a bit different !
We compared the new and the old version on a NT/PL statement on a 100 Mbyte ntuple plotting both a simple ntuple variable and a mathematical expression computed over the same variable.
The test has been repeated booking the resulting histogram first, to figure out the overhead due to the histogram min/max computation. The macro file is something like that:
hi/fil 1 big-charm.hbook * A fake plot just to fill the caches nt/pl 10.px4 * Book two histograms 1d 11 'Simple var' 100 -50 350 1d 22 'Simple expr' 100 -1 1 * NT/PL using the default resulting histogram nt/pl 10.px4 nt/pl 10.sin(px4) * NT/PL using previously booked histograms nt/pl 10.px4 idh=11 nt/pl 10.sin(px4) idh=22 * close 1
Results are summarized in the following table:
| OLD version | NEW version | with IDH ? | |
|---|---|---|---|
| Simple variable |
2.6 s |
10.6 s |
NO |
| Simple variable |
2.6 s |
5.5 s |
YES |
| Expression |
7.5 s |
13.8 s |
NO |
| Expression |
7.5 s |
7.0 s |
YES |
We see that the use of a predefined histogram gives a speedup by factor 2 on the new PAW, which looks logical since when no histogram is previously booked, data have to be scanned twice, one for the min/max computation, one for the histogram filling.
It seems that the old PAW does not have a penalty for min/max computation . The comparison of 'Simple variable' VS. 'Expressions' is interesting as well: while the old PAW is much faster when plotting a simple ntuple variable, the new version is a bit faster when a function has to be plot (which is the most realistic case).
 |
Release Notes | Known bugs | FAQs | Contributions | Tutorial | Reference manual | Down load | Miscellaneous |