Run multiple times and take mean speed Present results in a table for easier comparison Add some tricks to avoid optimization reordering stuff that breaks the benchmark
Consistency with rest of boost
Almost all of boost does, so be nice stay consistent