I see. Yes, that sounds like a great idea. How well does Immix cope with
high allocation rates of short-lived objects? Been a while since I read the
Immix paper...

Cheers,
Jon.

> -----Original Message-----
> From: caml-list-bounces@yquem.inria.fr [mailto:caml-list-
> bounces@yquem.inria.fr] On Behalf Of Benedikt Meurer
> Sent: 28 November 2010 20:00
> To: caml-list@yquem.inria.fr
> Subject: Re: [Caml-list] OCaml GC [was Is OCaml fast?]
> 
> 
> On Nov 28, 2010, at 20:40 , Jon Harrop wrote:
> 
> > I don't understand why this would help here though. Wouldn't that
> help when a long-lived structure was single large block but, in this
> case, the long-lived structure is a tree composed of many small heap-
> allocated blocks and, therefore, they would undergo the same wasteful
> "allocate in young only to be copied to old" pathological behaviour?
> 
> There is no "young" and no "old" in this scheme. There are two
> different heaps, one for large objects, one for small (probably 2-8k
> max object size, so not really small in the sense of OCaml). The small
> area is managed by Immix, which avoids copying of long-lived objects if
> they are allocated together with other long-lived objects (likely), or
> if not evacuates a set of probably related objects to a single chunk
> (this is somewhat dependent on the evacuation strategy, which will be
> differnt for OCaml compared to Java), further improving locality. There
> are simple heuristics to ensure that an object is not evacuated too
> often (it is already unlikely with the base algorithm), so there will
> be a lot less copying. One difficulty remains however: the pause times.
> It would probably be necessary to adopt Immix to a (semi-)incremental
> style to get the same pause times as with the current GC.
> 
> Benedikt
> _______________________________________________
> Caml-list mailing list. Subscription management:
> http://yquem.inria.fr/cgi-bin/mailman/listinfo/caml-list
> Archives: http://caml.inria.fr
> Beginner's list: http://groups.yahoo.com/group/ocaml_beginners
> Bug reports: http://caml.inria.fr/bin/caml-bugs