avoid inexact iteration count in benchmarks