Geometric mean
The geometric mean of n positive values is the nth root of their product.
For positive x_i, G = exp(sum(ln(x_i))/n). Declare any weighting and ratio orientation. SPEC CPU2026 uses geometric aggregation of specified benchmark ratios; that is a suite scoring rule, not a measurement of one combined storage workload. [1] [3]
With speedup defined as reference runtime / test runtime, constructed reference times [100, 100] s and test times [50, 200] s give ratios [2, 0.5] and geometric score 1. Executing each job once takes 200 s versus 250 s, so total-work speedup is 0.8. A geometric score of 1 does not imply equal total runtime or throughput. [2] [4]
Use the same workload set and weights when comparing normalized scores. Specify failures, timeouts and zero-handling separately: a failed run is not automatically a valid positive ratio, and silently removing it changes the population. The geometric mean does not recover raw elapsed time or transferred byte counts that were not retained. [2] [3]
Relations
- related
- Harmonic mean, Microbenchmark
Citations
- NIST Dataplot: Geometric Mean, product definition
- Fleming and Wallace, How not to lie with statistics: the correct way to summarize benchmark results, CACM 29(3), 218-221 (1986), rules 2-3 and Table IX
- SPEC CPU2026 run rules, sections 4.10.1-4.10.2: individual ratios and geometric aggregation
- Hesela: exact synthetic rate and benchmark-ratio examples, executable module