Moving discussion to vm-dev for follow up

On Sun, Jul 24, 2011 at 01:25:37PM +0200, Andreas Raab wrote:

On 7/24/2011 10:11, Hilaire Fernandes wrote:

...

Looking at the profile, I see in the iPad a lot of time spent in the method #idleProcess. The #eventLoop also counts for a lot. I don't know how to interpret these two results.

If time is spent in the idleProcess it is usually a good sign. Time assigned to the idleProcess is time in which your system didn't do a thing, so you have resources to spare. Usually this means you were running a UI interaction in which case there is always some idle time inbetween the Morphic interaction loop.

The bad news about the profile is the time spent in the eventLoop. The reason it's bad is that this is a clear sign that the profiler is bogus. I have commented on that many times in the past; basically the profile you have is almost entirely useless (see [1] for a good example of how bogus the results are).

Your best bet at this point is probably to go back to a VM that doesn't have the problem (probably as far as 3.6 if that's feasible) or to extract the OpenQwaq system profiler which you could then run on the desktop (I'm not sure if it's supported by the Stack VM; you could try it to see if it works on the iPad as well).

[1] http://lists.squeakfoundation.org/pipermail/vm-dev/2007-June/001410.html

This discussion seems to have died out prematurely without conclusion.

I opened Mantis 7653 to document the issue:

http://bugs.squeak.org/view.php?id=7653

My impression from reading the earlier discussion is that MessageTally accuracy was sacrificed in an attempt to gain performance, and that the right thing to do is to add the interrupt checks as suggested by Andreas, measure the actual impact on performance, and leave the interrupt checks in place if performance does not measurably suffer.

I cannot follow up further right now, but hopefully the Mantis report will keep the issue from being lost (again).

Dave