Performance engineering · 2026
P99KV
A C++20 and Linux performance lab built around an in-memory key-value store. The store is deliberately small. The measurement around it is the point.
Solo · open source, MIT
snapshot
- type
- Performance engineering
- period
- 2026
- source
- public github
- write-up
- A Timer Is Not a Measurement
problem
A benchmark that is easy to write is easy to believe. This one prints throughput and per-operation latency from the same workload, and the two disagreed with each other. The project exists to find that kind of thing: to learn what a change actually costs rather than what it ought to cost, and to distrust the instrument before trusting the number.
outcomes
- A 59% gain in GET throughput from removing an allocating copy, across five interleaved runs at -O2 on a pinned core, with the before and after ranges not overlapping on any run
- The per-operation latency column was found to be dominated by clock overhead, and is documented as such rather than reported as a result
- Published with the method, the hardware, the compiler and the commit, so the numbers can be re-run rather than taken on trust
what i built
- Deterministic workload: 50,000 entries, 32-byte values, 200,000 operations, and a fixed-seed access pattern so every run touches the same keys in the same order
- Throughput and latency measured in separate loops, with the throughput loop containing no clock calls at all
- 10,000 untimed warm-up operations before each measurement, and a printed checksum so the optimiser cannot delete the work being measured
- A separate single-workload binary, so one operation type can be run on its own under perf
- A build that defaults to -O2 rather than leaving an unoptimised benchmark as the obvious path, and a README that records the conditions behind the published numbers
tech stack
C++20CMakeLinuxperfg++WSL2
Read the source
The repo has the code, structure, and implementation detail.