Rather than running complete round trips for each pass, I ran the output generation passes first and then the input processing passes, so that I could measure the times separately.