[Python-Dev] binary trees. Review obmalloc.c

["Martin v. Löwis"]

Vladimir 'Yu' Stepanov wrote:
> * To adapt allocation of blocks of memory with other alignment. Now 
> alignment is rigidly set on 8 bytes. As a variant, it is possible to 
> use alignment on 4 bytes. And this value can be set at start of the 
> interpreter through arguments/variable environments/etc. At testing 
> with alignment on 4 or 8 bytes difference in speed of work was not 
> appreciable.

That depends on the hardware you use, of course. Some architectures
absolutely cannot stand mis-aligned accesses, and programs will just
crash if they try to perform such accesses.

So Python should err on the safe side, and only use a smaller alignment
when it is known not to hurt.

OTOH, I don't see the *advantage* in reducing the alignment.

> * To expand the maximal size which can be allocated by means of the 
> given module. Now the maximal size is 256 bytes.

Right. This is somewhat deliberate, though; the expectation is that
fragmentation will increase dramatically if even larger size classes
are supported.

> * At the size of the allocated memory close to maximal, the
> allocation of blocks becomes inefficient. For example, for the
> sizesof blocks 248 and 256 (blocksize), values of quantity of
> elements (PAGESIZE/blocksize) it is equal 16. I.e. it would be
> possible to use for the sizes of the block 248 same page, as for the
> size of the block 256.

Good idea; that shouldn't be too difficult to implement: for sizes above
215, pools need to be spaced apart only at 16 bytes.

Regards,
Martin