eapache / wof_alloc Goto Github PK

In practical usage in Wireshark, it appears that the hotspot in the allocator is the split_free_chunk method. This method has already been relatively optimized, but there is one potential optimization remaining.

The free-chunk header is currently stored at the front of the chunk. This means split_free_chunk has to do a bit of song and dance to move it when splitting the chunk, especially if the allocation is small and the new chunk might overlap the header. Placing the free-chunk header at the end of the chunk instead would simplify this and speed up split_free_chunk, however it would add certain other costs which may or may not outweigh the benefits:

• Accessing the free header from the chunk header would require one additional subtraction.
• The free header and the chunk header would no longer be next to each other, so cache hits might go down.

Need to write this and then benchmark it. The callgrind trace at https://bugs.wireshark.org/bugzilla/show_bug.cgi?id=9823 indicates a good test-case which hits this hotspot.

Copy from https://code.wireshark.org/review/492 once it is merged and tested.