In the past I've sometimes achieved 20% or better speedups by writing custom allocators (speedup on overall performance, not just malloc performance). tcmalloc or jemalloc are great for the general case, but sometimes you know invariants about object sizes, alloc patterns and free patterns that allow much more performant allocation.

The simplest case is if you know that you will free everything at once, or nothing at all. This allows you to eliminate most bookkeeping and allows a completely lock-free architecture. But there are also more complex cases where you can still get big benefits from exploiting known invariants.