Performance (and sensitivity) studies in parallel job scheduling mostly use average values of the measurement metrics over the entire job stream. This does not give an idea of relative job performance (hence starvation) ...