Boost logo

Boost :

Subject: [boost] Preventing Boost.Pool from allocating too much memory - need your help.
From: John Maddock (jz.maddock_at_[hidden])
Date: 2011-02-16 13:28:38


There is a long standing bug against Boost.Pool that points out that it
allocates far too much memory when the size of the requested block is oddly
sized: https://svn.boost.org/trac/boost/ticket/1252

For example if you request blocks of size 1501 bytes then you actually get
6004 byte sized blocks because Pool uses LCM to calculate a size that can
store both a pointer and the requested block.

There's some dispute on the Trac ticket whether the correct solution is to
just round up the block size to a multiple of the pointer size. I'd like to
run my take on this past you all and check I have the right fix ;-) !

So here goes - if we round up to the nearest multiple of
pointer-alignment-size then:

1) The actual alignment requirements of the block (which only the client of
Boost.Pool knows) must be a factor of the block size (otherwise arrays of
blocks would be incorrectly aligned).
2) If the requested block has an alignment requirement >= pointer type, then
of necessity alignment_of<void*>::value will be a factor of the block size
and no rounding occurs or is necessary.
3) If the requested block has an alignment requirement < pointer type, then
we may round up the block size, but this will only *increase* the alignment
of the returned blocks so is always safe.

Note that (2) assumes that all alignments are powers of 2, if that isn't
true, then all bets are off. But I'm assuming there are no real-world
architectures that violate that assumption?

Thanks in advance, John.


Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk