[Haskell-cafe] Re: Real-time garbage collection for Haskell

[Simon Marlow]

On 01/03/2010 14:16, Sönke Hahn wrote:
> On Monday 01 March 2010 01:04:37 am Luke Palmer wrote:
>> On Sun, Feb 28, 2010 at 2:06 AM, Pavel Perikov<perikov at gmail.com>  wrote:
>>> Did you really seen 100ms pauses?! I never did extensive research on this
>>> but my numbers are rather in microseconds range (below 1ms). What causes
>>> such a long garbage collection? Lots of allocated and long-living
>>> objects?
>>
>> This is all hypothetical right now.  I heard some horror stories in
>> which people had to switch to the main game loop in C++ and only do
>> the AI logic in Haskell because of pauses.  I would rather not do
>> that, especially because this project is *about* proving Haskell as a
>> viable game development platform.  So I am trying to be prepared if I
>> do see something like that, so that it doesn't put the show on hold
>> for a few months.
>>
>> Presumably large, long-living objects would cause the generation 0
>> collections to take a long time.  I am not sure if we will have any
>> said objects, but we can't rule it out...
>>
>> Thanks for the positive reassurances, at least.  I'd like to hear from
>> people with the opposite experience, if there are any.
>
> Yes there are. I am working on a game with Haskell using OpenGL rendering.
> I've done some frame time measurements lately and encountered single frames
> needing more than 100ms to be rendered. I am currently trying to gather
> information on what is going on in these 100ms and why. From what i
> understand, the GC is running very often and just some (very few) of its runs
> are very slow.
>
> BTW: switching to parallel GC (either with -N1 or -N2 (on a dual core
> machine)) made the behavior MUCH worse, for some reason.

Probably due to cache effects - shipping the data off to a different 
core in the GC can far outweigh any benefits you would have got by 
traversing the heap in parallel.  For a single-threaded program, 
parallel GC tends to only help when the heap gets large (over 50MB at a 
guess).  For parallel programs, parallel GC helps by keeping the data in 
the right cache.

I'm finding that for parallel programs turning off load-balancing with 
+RTS -qb often helps, because load-balancing is bad for locality. 
Again, if the heap gets large, then collecting it in parallel is 
probably more important than keeping it in the cache.

Cheers,
	Simon