Em 16/01/2010 18:00, John M McIntosh <johnmci@smalltalkconsulting.com> escreveu: John,
On 2010-01-16, at 9:11 AM, csrabak@bol.com.br wrote:
I agree with the intent but cannot figure out a means for evaluating (testing) the tweaking of the parameters in 'scientific' way!!
Well some clues are in http://www.smalltalkconsulting.com/papers/GCPaper/GCTalk%202001.htm See the "Squeak Memory Layout" page
Is there a way to download a file of the presentation (be it powerpoint or PDF)?
But really you have to have an understanding of how the GC behaves, and then the various parms (which aren't in the 2001 paper).
OK. Is there other reference?
then some Sunits, and generate charts like slide 92, then decide how to alter behaviour.
<thinking aloud> Given the multitude of platforms and configurations, perhaps instead (or in addition to) SUnit tests, we should create some script or some code that runs and do a diagnostic suggesting the best tuning for the user. </thinking aloud>
Having Sunits is a important criteria since it then tells you if changes are impacting time to run in a meaningful way, since altering the values can have a negative effect.
Yes. Some measurable and repeatable way to do it is fundamental in order not to jeopardize the performance of Pharo. The effort to revert bad first impressions would Herculean! -- Cesar Rabak