Re: Vectors/Atomic Tuples/Garbage Collection

[Eero Pajarre <epajarre@...>]

Dylan Cuthbert wrote:

> > Also I do avoid creating tables (avoid, but not completly disallow)
> > during normal operation, I have been able to keep my garbage generation
> > relatively modest so far.
>
>
> It seems like you're trying to fight a tsunami with a feather as the
> complexity and number of your tasks increase though.  The problem is, unless
> the garbage collector is very smart its going to have to go through pretty
> much every single object in existance checking its reference count, when a
> few quick checks after an expression would solve the problem.  Going through
> every object at the end of every frame seems to be a major amount of
> overhead.  Of course, the garbage collector might not be as naive as this, I
> haven't analysed that part of Lua yet.

This of course depends among other things on the division between
Lua and the C++ (in my case) world. I admit that most of my data
is in the C++ side. On Lua side I only have couple of tables for
each "intelligent" object for my world.

As it is, I am doing so much computation in Lua that the interpreter
speed is an issue. Surprisingly the GC is not significantly visible
in my profiling results (Vtune & AMD Codeanalyst). My worst
CPU spikes happen in Lua internal table search (apparently VC++
manages to create code which doesn't suite my processor well), and
in the lua virtual machine code, which really breaks branch
prediction.

It still seems that I am lucky enough, that I can implement my
project to the end with current Lua GC behaviour if needed, even
if I force full GC 50-75 times per second.

			Eero