Thanks for pointing out the article, made for interesting reading...
but I'm afraid I have to say your followup comparisons are unfair...
>One interesting point is that TYA is on various benchmarks
>2-8 times slower than the Symantec jit included with Sun's java
hmmm... one is developed by a large international company over the
course of a couple years, by an estimated few dozen engineers and costs
money (even if _you_ got it for free, SUN paid Symantic for it); the
other is developed by a small handfull of developers in spare time as
OSS and costs nothing. The one with jdk116+ from Sun is the "little
sister" of the commercial JITC Symantic sells -- reported to be the
best in the industry. Also look at the interpreted only numbers and
you'll see that the win32 jvm is a tad faster than the Linux jvm to
begin with. That blackdown/TYA even _made_ the short list to be
compared is a tribute to the great work by everyone on the two teams!
This was mostly a Win32 article, with a pair of Linux JVMs thrown in,
probably because of Tower's touted performance. Note the absence of:
OS/2, OS/400, OS/390, AIX, HP/UX, SCO, Reliant, Solaris, Be and Mac
in the platform catagory and Kaffe and (whatever that other one is) in
the Linux JVM catagory, and the projects such as Cygnus' GCJ or IBM's
High-Performance compiler for Java in the "let's bend the JVM concept
as far as we can in the name of performance" catagory (that's where I
put TowerJ). Also missing were the other linux jits.
>under Windows (which in turn is slower than the forthcoming hotspot).
hotspot isn't a JIT, it's a whole new breed of VM...
>Since TYA is roughly the same speed as the 'sunwjit'
>jit supplied by sun for the linux java1.2, this means that Sun supplied
>linux a much slower jit than the one it supplies with Windows.
HUH?! I don't understand this logic... are you comparing 1.1 to 1.2???
or symantic with sunwjit??? either way I gotta say "no fair"....
>Is this because they cannot legally supply the Windows jit? Where
>did this jit come from?
Which windows jit? the Symantic jit they shipped with 116+ or the "sunwjit"
they ship with 1.2? the first one they licensed binaries from Symantic,
originally in the "preformance pack" they had for 115 (or maybe it was 114)
the second I'm told was developed in house....
[timings snipped]
very interesting numbers... but any benchmark as short as those really
aren't very good if you only want to look at the java time... you also
captured program load, kernel overhead, etc... which in that short of
time is probably a substantial percentage. a good way to avoid that is
to capture System.currentTimeMillis() before and after a "test" and
subtract... an even better mark is to loop the test a few hundred
times and divide the differences by that mark... for example, if I
wanted to benchmark the performance of my render() method...
long before, after;
static final int loops = 500; //incr for quicker methods
for (int lcv=10;lcv>0;--lcv) {
render();
} //warmed-up the jit, etc...
System.gc();
System.out.println("warmed up");
before = System.currentTimeMillis();
for (int lcv=loops;lcv>0;--lcv) {
render();
// System.out.println("-=- "+lcv+" -=-"); //use only when testing for gc!
}
after = System.currentTimeMillis();
System.out.println("avg:"+( (after-before)/loops )+"ms.");
then run with -verbosegc and try to adjust -ms and -mx (both the same)
to a value so that you minimize GC, without paging... you can use fewer
loops at that point... from the times you posted for splinefill, I'd
use about 10,000 loops; and for random about 400. -=Chris
p.s. (_way_ off-topic) /pix/sunset640sig.jpg is gorgeous, as is the
background of /pix/cvg640.jpg (not sure what the object is though...)
!NEW!-=> <*> cabbey at home dot net http://members.home.net/cabbey/ <*>
"What can Microsoft do? They certainly can't program around us." - Linus
-----BEGIN GEEK CODE BLOCK----- Version:3.12 http://www.geekcode.com
GCS$/IT/PA$ d(-) s++:+ a-- C+++$ UL++++ UA++$ P++ L++ E- W++ N+ o? K? !P
w---(+)$ O- M-- V-- Y+ PGP+ t--- 5++ X+ R tv b+ DI+++ D G e++ h(+) r@ y?
------END GEEK CODE BLOCK------
----------------------------------------------------------------------
To UNSUBSCRIBE, email to [EMAIL PROTECTED]
with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]
