I am trying to replace an old implementation of a small object memory allocator with boost::pool. I have noticed that boost::pool uses quite a bit more memory for certain requested_size than the allocator I am trying to replace.

After looking at the code and reading the Guaranteeing Alignment document, I am wondering why boost::pool::alloc_size() have to be lcm(sizeof(void*), sizeof(size_type), requested_size)? Lots of bytes can be wasted in the chunks returned to user when requested_size is not a multiple of min_alloc_size (lcm(sizeof(void*), sizeof(size_type))).

I would think the requested_size alignment could be the first multiple of min_alloc_size that is larger than requested_size. For example, the convoluted example in the alignment document has requested_size==7, sizeof(void*)==3, and sizeof(size_type)==5. Currently alloc_size() returns lcm(3,5,7)==105, but since min_alloc_size is lcm(3,5)==15, I think requested_size alignment could be 15 bytes, rather than the 105 bytes.

Another way to look at this. If requested_size was changed to 15, then the current implementation of alloc_size() would return 15 bytes (lcm(3,5,15)==15), so why the huge penalty when requested size is 7?

Am I missing something?

Thank you.

Soren