[MLton-devel] D2 problem

[Stephen Weeks]

> The  thing  worrying  me  about  GC  slowing down with multiple chunks is any
> things like  tests  for  from/to  space,  which  now  are  considerably  more
> expensive. 

I don't think we ever need to do that kind of test except in debugging
code.

> Also some allocations will fail (because of chunk full) but will
> not trigger a GC, but instead put the data in a  different  chunk.   If  this
> happens much (10 meg chunks and 1 gig of semi-space would mean only 100 times
> per GC) then it will slow things down.

I think with big enough chunks we should be OK.  Another possibility
is to use the smaller chunks as a fall-back.  That is, always try to
allocate the semispace as we would have, but only if we fail do we use
smaller chunks.

> As  to  the mmap under Cygwin stuff, I'm remembering Linux 2.2 (or maybe 2.0)
> where by specifying a non-zero first arg you would cause the  kernel  to  get
> the  space  from  a different part of your address space, and that this would
> cause you to be able to get more  memory.   The  same  could  be  true  under
> Cygwin,  so that there is a way to get more than 2 gig, but just not 2 (or 1)
> continuous gig.

The 2G limit is a Win NT/2000 limit.  I don't think it's possible to
work around that.  The 1.1G limit is due to address space
fragmentation and may be worked aroud.

> You  didn't  answer how important it is to the Polyspace people to be able to
> run these big (> 2 gig combined spaces)  programs  under  Windows.   Sadly  I
> would guess rather important.

We'll see what Alain says when he gets in on Wednesday.

_______________________________________________
MLton-devel mailing list
MLton-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/mlton-devel