[Mono-dev] [PATCH] Block map support for sgen

[Miguel de Icaza]

Hello,

   For the sake of us that do not really speak the GC lingo, would you
mind explaining what block map support for SGen is?

> 
> The attached patch set implements block map support for sgen. It uses
> a schema similar to boehm's, which is a 2 level sparse map.
> Under 64 bits it uses hashing.
> 
> 
> I benchmarked a modified binary-trees without valuetypes. Block maps
> gives a very modest speedup under major-copying (about 2%) and
> nothing under major-marksweep. I've only used the block map for
> major_copy_or_mark_object thou. There are probably other places it
> oould be used too.
> 
> 
> The design is basically the same as boehm's except for a few things:
> 
> 
> -It doesn't store list heads or address on each segment. This allows
> segment's size to be a power of 2; and
> -LOS is handled by filling all covering slots with its block instead
> of using forwarding
> 
> 
> Few notes:
> 
> 
> Segments are not deallocated since this requires either scanning whole
> segments on each deallocation or keeping block counts.
> And it's probably not needed since Boehm doesn't do it. It's doable as
> long as the block map is only read during GC and mutated
> with the gc lock held.
> 
> 
> 64bits support has not been committed since it is a minor change to
> the code in sgen-gc.c and I want to have the current change set
> validated first.
> 
> 
> A small config option that uses either a 3 level map or just hashing
> under 32bits can be done with ease.
> 
> 
> The embedding of Block in MSBlockInfo wastes a word of memory. This
> could be worked out by either factoring Block::role into a separate
> struct or by using Block::next in place of MSBlockInfo::next.
> 
> 
> Cheers,
> Rodrigo
> _______________________________________________
> Mono-devel-list mailing list
> Mono-devel-list at lists.ximian.com
> http://lists.ximian.com/mailman/listinfo/mono-devel-list