Hello all, How would people feel about an alternate GC implementation in GHC which is not parallel? A GC like this would be simpler to understand, maybe a little faster when parallel collection is not being used, and (most importantly for my case) easier to extend with interesting features. In particular, this implementation would not have gen_workspace or gc_thread; these would either be unnecessary or folded directly into the actual generation object. Thanks, Edward
On 10/08/13 00:58, Edward Z. Yang wrote:
How would people feel about an alternate GC implementation in GHC which is not parallel? A GC like this would be simpler to understand, maybe a little faster when parallel collection is not being used, and (most importantly for my case) easier to extend with interesting features.
In particular, this implementation would not have gen_workspace or gc_thread; these would either be unnecessary or folded directly into the actual generation object.
From my point of view I'd like such a thing to be as separate as possible from the rest of the GC code. However it might be hard to do that - you mentioned modifying the generation structure, for example. As with most things in the RTS, the GC is "exactly as modular as it needs to be right now", which in practice means "not quite modular enough for what I want to do next" :-) So your first step might be to abstract some things so that the two GCs can coexist peacefully. But perhaps there's another way to achieve your goals - what are the interesting features you want to add? Cheers, Simon
But perhaps there's another way to achieve your goals - what are the interesting features you want to add?
Yeah. The key thing I need to change is how the GC decides where live objects are evacuated to, to support a more efficient implementation of resource limits (think BiBoP for cost centers) where every user gets his own set of pages, and his objects are always evacuated to pages he owns. I don't know how to parallelize his, and even in the non-parallel case it requires quite a restructuring of the GC code. Edward
On 13/08/13 14:15, Edward Z. Yang wrote:
But perhaps there's another way to achieve your goals - what are the interesting features you want to add?
Yeah. The key thing I need to change is how the GC decides where live objects are evacuated to, to support a more efficient implementation of resource limits (think BiBoP for cost centers) where every user gets his own set of pages, and his objects are always evacuated to pages he owns. I don't know how to parallelize his, and even in the non-parallel case it requires quite a restructuring of the GC code.
You can think of this as abstracting two operations: - deciding where to move the object - deciding whether that creates a cross-generation pointer (and if so, adding the parent object to the remembered set) I imagine this is independent of generational GC (each generation is split into multiple users) so the second question is unchanged - it just compares the generation numbers of the source object and the destination. So then you just need to manage the new sets of areas. Parallelism doesn't add much complexity, you just have a set of destination areas per generation per thread. Cheers, Simon
participants (2)
-
Edward Z. Yang -
Simon Marlow