The runtime speed comparison feature seems to highly depend on the
system load and scheduler.

Maybe use another metric, such as the number of instruction executed,
or the sum sys + user time (instead of real time) might give a better
performance indicator (the linux kernel provides all the tools to measure those value).