I like taking the trimmed mean of 10-20 runs, or if a run is quick, the (half-sample) mode of more runs. See robust_statistics.h.