Hi all,
I've been working with some large (~2000 line) and sophisticated stylesheets
to generate svg. Now that things in xsltc appear to work for us I'd like to
share some performance figures.
Most timings were done at the win2k command-line level - as such, library
load time should be carefully considered (in this vein, stylesheet 0 is an
identity transform on an empty document).
The timing labeled 'avg' was made with a patched up version of the
cmdline.Transform class such that the exe load, the document load and the
1st translet run happen outside the timing loop (the 1st translet run does
all kinds of introspection on the classpath). I suppose this represents a
best-case performance figure.
|----translet-------|
stylesheet saxon xt compile run avg note group
---------------------------------------------------
0 44 37 322 108 $
1 69 57 435 134 1 a
2 1712 796 570 2252 1938 ? b
3 201 131 561 281 74 ? a
4 148 92 561 247 62 ! a
5 244 130 563 279 69 ? a
6 75 58 445 121 2 a
7 166 117 560 234 54 ? a
8 163 96 585 259 72 ? a
9 152 102 594 234 55 ! a
10 610 733 585 796 609 * c
11 132 94 532 235 48 * c
12 166 106 559 243 68 *? c
all times in hundeths-of-a-second
tests done using win2k with jdk 1.3.1.
run = library load + 1 doc load + 1 translet run
avg = avg of: 5 translet runs (less library, doc and 1st run)
$ = identity transform on empty document
* = translet trig emulation
? = minor fidelity problems
! = major fidelity problems
The timings in column 'avg' and would appear to represent an average speed
improvment of 98/49=100% over xt for the 'group a' stylers (group b is all
wacked-out, and group c is running with trig emulation so these groups were
not included in the calculation).
Since some of the results (like for stylesheet 2 and 10) look funny I'll be
focusing my attention in their neighbourhood next.
Pleasure working with all of you,
john
