2011/2/10 Felipe Ba�ados Schwerter <fbanados@dcc.uchile.cl>

Hi Everybody! This is my first post on the list, even though i've been subscribed for more than 6 months.

2011/2/10 Levente Uzonyi <leves@elte.hu>

On Thu, 10 Feb 2011, Torsten Bergmann wrote:

see http://fbanados.wordpress.com/2011/02/10/a-tinybenchmark

Maybe Eliot is able to say something here on the main
differences between Cog <-> VW VM.

You made several mistakes:
- You didn't post the source code for #tinyBenchmarks. Do all those Smalltalks implement it? Are the implementations equal?

No, they don't implement it. Just Pharo and Squeak. I just added the same implementation to Integer in VA, VisualWorks and GNU Smalltalk ( Integer>>tinyBenchmarks, Integer>>benchmark and Integer>>benchFib), so the source code is the same.

- You didn't specify the version numbers for the VMs used. Or if you built them yourself, then the sourcecode + compiler version + configuration parameters. Without these your measurements can't be reproduced.

Every System was the one available from their websites. (Got Pharo 1.2 from Hudson about 3 days ago.) �However, VisualWorks is not available from its website right now. I installed it from a Cincom CD-ROM I got on Smalltalks 2010 .

�
- There are two different Cog VMs. You didn't specify which one did you use with Pharo.

Smalltalk vmVersion returns�'Croquet Closure Cog VM [CoInterpreter VMMaker-oscog.35]').

But Smalltalk getSystemAttribute: 1008 returns the JIT version, which is much more important for this performance measurement.

�

�
- You used a CogVM (JIT) with Pharo but SqueakVM (interpreter) with Squeak 4.2. What's even worse is that you based your conclusion on this mistake.

Yes, it is true. Daniel Galdames also pointed this out last night on the blog. But as I told before, it is just the benchmarks you get from the "download the latest version available" approach. I couldn't find a Cog-based Squeak on its ftp site, (perhaps I was too lazy). I won't have the time to repeat this until next week, then I'll try this.

�
- Does #tinyBenchmarks alone give fair performance comparison? Hardly.

Well, it was the benchmark I knew of. Do you know a better performance benchmark available? If so, I could retry the experiment.

As large a benchmark that is as close to your working case as possible. �tinyBenchmarks is little more than a smoke test (which is what we use it for at work; if tinyBenchmarks hasn't plummeted ten the build is OK performance-wise).

In the Cog-Benchmarks (in Cog package) that I uploaded to squeaksource recently you'll find ShootoutTests which has 4 tests from the computer language shootouts that are pure Smalltalk.

�

�

Levente

Thx
T.
--
NEU: FreePhone - kostenlos mobil telefonieren und surfen!
Jetzt informieren: http://www.gmx.net/de/go/freephone

--
Felipe Ba�ados Schwerter
Estudiante de Ingenier�a Civil en Computaci�n
http://www.dcc.uchile.cl/~fbanados
Universidad de Chile