Context

The BSON encode/decode perf benchmarks are too noisy, making a small regression hard to distinguish. Network is not a factor for these tests, so we should be able to reduce the noise.

While working on abi3 support for the driver and benchmarking on a perf host, I identified the following potential improvements.

• Pin each benchmark to a physical core using os.sched_setaffinity (when available on the OS).
• Use time.process_time_ns instead of time.monotonic.
• Warm up before timing and discard the first ~20% of iterations.

Definition of done

• Validate and implement the suggested approach(es)

Pitfalls

• Run the benchmarks on 10-20 separate runs to compute the baseline in terms of Coefficient of Variant (CoV), computed as std dev / mean
• For a given improvement, run the same number of times and compare the CoV.
• Only keep a change only if it brings CoV to ≤1%, or reduces it by ~25% or more, otherwise it isn't worth the change.