[v3] allocator documentation revamp

[Felix Yen]

Benjamin, I'm flattered by the reference, and coincidentally, just 
updated yalloc earlier today.  (Support for reconfiguration was broken, 
and it now implements Bonwick's coloring technique.)  However, the 
yalloc URL has a half-life of 1-2 months and I haven't decided if that 
code will remain on the 'net afterward.  There isn't enough interest in 
it to justify adding it to SourceForge.

Stefan, I think removing the reference is probably the best solution.  
In case you're wondering, yalloc differs from mt_allocator primarily in 
three ways.  First, yalloc doesn't use internal thread ids because I 
wanted to support VC++ 6 and didn't want to use a third-party 
implementation of Pthreads.  After encountering several conformance 
problems with that compiler and seeing how these ids contribute to 
mt_allocator, I regret that decision as it produced yalloc's most 
significant limitation.

Secondly, yalloc uses one bin per object size.  This is primarily 
because it is supposed to be used very sparingly, e.g. for fewer than 
ten classes/containers in large systems.  (This intended usage also 
justifies using one mutex for the entire allocator and ignoring false 
sharing.)  Rounding object size to the next power of two is appropriate 
for a default allocator, but it can create locality of reference 
problems for the larger bins, as unrelated containers allocate memory 
from the same free list.  Even yalloc is impacted by this issue, but to 
a lesser extent.  (Nathan mentions the problem in his 1993 article.)

Thirdly, yalloc recycles entire free lists instead of blocks.

The minor differences include yalloc's introducing a linker dependency, 
its use of coloring, and its support for overriding new/delete.


Felix