It might sound surprising, but slowing down programs has various use cases: it helps exposing race conditions, facilitates virtual speedup estimation, and allows one to assess a profiler’s accuracy. Yet, slowing down a program is complicated because today’s CPUs and some runtime systems optimize execution on the fly.

While the basic requirement for the mentioned use cases is to preserve a program’s correctness, the more challenging requirement is to preserve a program’s performance behavior to avoid introducing unintended bias.

In this work, we evaluate six instruction candidates to achieve controlled and fine-grained slowdown without biasing performance, including e.g. NOP, MOV, and PAUSE. We tested each candidate’s ability to achieve a target overhead of 100%, whether the profiler-observable performance behavior changes, and whether slowdown placement inside basic blocks influences results.

On an Intel Core i5-10600 with the Comet Lake-S microarchitecture, our experiments suggest that only NOP and MOV instructions are suitable. We tested this in the context of the Graal just-in-time compiler, with a representative benchmark that has 53 basic blocks. Based on previous work, we believe these experiments can guide future research on advanced developer tooling that utilizes fine-granular slowdown at the machine-code level.