Ned Konz <ned at bike-nomad.com> wrote:

> On Sunday 28 March 2004 8:16 pm, Tim Rowledge wrote:
> > I wonder how far we are? I wouldn't be surprised if we're better than
> > 2x anyway simply from general improvements in the code.
> 
> As you may recall from your visit to my house, I've seen a 2:1 spread over 
> various *3.6* VMs on my system, depending on optimization flags and GCC 
> versions.
Hmm, forgot about that detail. It's certainly true that the newer CC on
RISC OS added a good 10%. Just adds more confusion to the issue.
> 
> > Wish my recent 
> > prim/timer changes had done as much good for x86 as they did for RISC
> > OS!
> >
> > I'd suggest doing benchmarking using the benchmark package on SM; it
> > includes the old GreenBook benchmarks, slopstones and smopstone and
> > ought to give a good spread of data. If someone actually has a 3.2vm
> > that can run an image that can load the benchmark code we could find
> > out the answer. I'd be interested, quite aside from money.
> 
> On a Pentium Pro system running Linux (2.4GHz, 800MHz memory bus), and on a 
> 3.2-4 image, I get:
> 
> 3.2-4 VM, Linux
> 0 tinyBenchmarks '126357354 bytecodes/sec; 3343528 sends/sec'
> Performance Rating		5141.30810938756
> 
> 3.6g-3 VM, Linux
> 0 tinyBenchmarks '207792207 bytecodes/sec; 6490934 sends/sec'
> Performance Rating		12107.95708688026
> 
> Details are attached.
Interesting; that's certainly >2x over all I'd say. I'm quite surprised
the micro-BMs vary so much as well as the macro. Perhaps the
global-struct change helpd there.

Guess it's time we had some real profiling of VM activity to ponder.


tim
--
Tim Rowledge, tim at sumeru.stanford.edu, http://sumeru.stanford.edu/tim
An algorithm must be seen to be believed.  - D. E. Knuth