Low-level array performance

[Dan Doel]

On Tuesday 17 June 2008, haskell at list.mightyreason.com wrote:
> I see that Dan Doel's post favoring Ptr/Addr#
> has the same allocation amounts (from +RTS -sstderr) for Ptr/Addr# and the
> MutableByteArray#
>
> Everyone else sees more allocation for Ptr/Addr# than MBA# and see MBA# as
> faster in these cases.
>
> I myself (on G4) see more allocation [just like Simon Marlow] for Ptr/Addr#
> and find it slower.  If I boost the initial memory with "-A 100m" then Ptr
> still allocated more, but the timing difference becomes quite small:

Pardon my noise, but is this still with the version of Ptr.hs that would (in 
your case) allocate a 1 million element list and traverse it twice, or the 
revision that fills the array in a loop with an Int#?

If it's the former, and Addr# is tying MutableByteArray# even with operations 
on a 40-some megabyte list (if the allocation is any indication), then the 
actual Addr# operations are probably faster for you, too. :)

I'll attach new, hopefully bug-free versions of the benchmark to this message.

Of course, without the list overhead, the ByteArr appears to allocate much 
more than Ptr for large arrays, because the n*w byte array shows up in the 
heap allocation, whereas the malloced memory does not. None of this should be 
a factor in the actual fannkuch benchmark, of course, which only allocates 3 
arrays of size 11.

Cheers,
-- Dan
-------------- next part --------------
A non-text attachment was scrubbed...
Name: ByteArr.hs
Type: text/x-hssrc
Size: 1407 bytes
Desc: not available
Url : http://www.haskell.org/pipermail/glasgow-haskell-users/attachments/20080617/e506727d/ByteArr.bin
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Ptr.hs
Type: text/x-hssrc
Size: 1079 bytes
Desc: not available
Url : http://www.haskell.org/pipermail/glasgow-haskell-users/attachments/20080617/e506727d/Ptr.bin