Parallel efficiency expresses speedup per thread. It helps show whether adding threads produces a proportional reduction in runtime. For speedup $S(p)$ with $p$ threads:
$$ E(p)=\frac{S(p)}{p} $$
If eight threads give a speedup of four, the efficiency is 0.5, or 50%. Ideal linear scaling gives an efficiency of one. This metric describes scaling relative to the serial baseline; it does not measure how busy each processor core is.
The benchmark plotting script calculates efficiency from each measurement's speedup and thread count, then plots the median for each configuration. Efficiency often falls as threads are added because synchronization and shared memory resources consume a larger share of runtime.
A configuration can be faster while having lower efficiency. For example, moving from four to eight threads may reduce runtime without doubling speedup. Choose according to whether elapsed time or economical use of cores matters more. Values above one are also possible in measurements, but require examining cache behavior, processor frequency, and variation before drawing conclusions.