Please sign in to comment.
mempool: significant reduction of memory waste
The mempool allocator implementation recursively breaks a memory block into 4 sub-blocks until it minimally fits the requested memory size. The size of each sub-blocks is rounded up to the next word boundary to preserve word alignment on the returned memory, and this is a problem. Let's consider max_sz = 2072 and n_max = 1. That's our level 0. At level 1, we get one level-0 block split in 4 sub-blocks whose size is WB_UP(2072 / 4) = 520. However 4 * 520 = 2080 so we must discard the 4th sub-block since it doesn't fit inside our 2072-byte parent block. We're down to 3 * 520 = 1560 bytes of usable memory. Our memory usage efficiency is now 1560 / 2072 = 75%. At level 2, we get 3 level-1 blocks, and each of them may be split in 4 sub-blocks whose size is WB_UP(520 / 4) = 132. But 4 * 132 = 528 so the 4th sub-block has to be discarded again. We're down to 9 * 132 = 1188 bytes of usable memory. Our memory usage efficiency is now 1188 / 2072 = 57%. At level 3, we get 9 level-2 blocks, each split into WB_UP(132 / 4) = 36 bytes. Again 4 * 36 = 144 so the 4th sub-block is discarded. We're down to 27 * 36 = 972 bytes of usable memory. Our memory usage efficiency is now 972 / 2072 = 47%. What should be done instead, is to round _down_ sub-block sizes not _up_. This way, sub-blocks still align to word boundaries, and they always fit within their parent block as the total size may no longer exceed the initial size. Using the same max_sz = 2072 would yield a memory usage efficiency of 99% at level 3, so let's demo a worst case 2044 instead. Level 1: 4 sub-blocks of WB_DN(2044 / 4) = 508 bytes. We're down to 4 * 508 = 2032 bytes of usable memory. Our memory usage efficiency is now 2032 / 2044 = 99%. Level 2: 4 * 4 sub-blocks of WB_DN(508 / 4) = 124 bytes. We're down to 16 * 124 = 1984 bytes of usable memory. Our memory usage efficiency is now 1984 / 2044 = 97%. Level 3: 16 * 4 sub-blocks of WB_DN(124 / 4) = 28 bytes. We're down to 64 * 28 = 1792 bytes of usable memory. Our memory usage efficiency is now 1792 / 2044 = 88%. Conclusion: if max_sz is a power of 2 then we get 100% efficiency at all levens in both cases. But if not, then the rounding-up method has a far worse degradation curve than the rounding-down method, wasting more than 50% of memory in some cases. So let's round sub-block sizes down rather than up, and remove block_fits() which purpose was to identify sub-blocks that didn't fit within their parent block and is now useless. Signed-off-by: Nicolas Pitre <firstname.lastname@example.org>
- Loading branch information...
Showing with 7 additions and 37 deletions.