A small, fixed, and purely synthetic workload used to calibrate the benchmark results.
Publishing its runtime alongside the real measurements allows a viewer to normalize the other numbers by it,
and hence to cancel out how fast or how loaded the machine was that produced them.
The workload must therefore stay deterministic (no file system, no R session, no randomness).
Changing the work or the round count puts every future run on a new scale. Whoever changes it has to
back-correct the published history by the shift, or say so in the name of the measurement.
A small, fixed, and purely synthetic workload used to calibrate the benchmark results.
Publishing its runtime alongside the real measurements allows a viewer to normalize the other numbers by it, and hence to cancel out how fast or how loaded the machine was that produced them. The workload must therefore stay deterministic (no file system, no R session, no randomness).
Changing the work or the round count puts every future run on a new scale. Whoever changes it has to back-correct the published history by the shift, or say so in the name of the measurement.