This seems like the kind of metric that 3 users with 15 year old machines can skew significantly.
Has to be normalized, and outliers eliminated in some consistent manner.
I'm pretty sure I saw them present on exactly this at FOSDEM?