I suspect a large portion of your performance improvement was just from code inlining. Your simple allocator was probably inlined, whereas malloc requires a function call (as the code lives in a shared library).