Let the scheduling latency for a thread be the time from the moment it is set runnable to the moment it is scheduled for execution. With an ad hoc patch we collected the maximum value for the scheduling latency at thread exit. These are histograms of the data and have no meaning whatsoever. Seriously.

The first set of data was collected on a 450MHz Pentium 3 during a kernel compilation:

This set of data was collected on a 2400MHz desktop:

Let's focus on the higher latencies. Before making this histogram latencies <10^-2sec were removed (p3):

Same here for the desktop: