Trying to pin down whether changes to the 2.5 and 2.6 scheduler have helped or hurt performance, especially on interactive programs, has been both difficult to quantify and very subjective. One favored test has been to create your favorite load and then move your cursor around and observe how slow or fast it is. Another one is to drag a window across your desktop and see how quickly it gets redrawn. And I would certainly be skewered if I didn't mention what is probably the favorite: playing your favorite music while under load and listening intently for skips.
Unfortunately, all these measurements are subjective, and even, at times, argumentative. With scheduler statistics installed, one can accurately measure such things as the amount of time processes are spending on the processor or the amount of time they are waiting for the processor. This means that on SMP and NUMA machines, load balancing efforts can be objectively evaluated, and process migration decisions more effectively reviewed. And all of this can be done with no measurable impact to the system.
This paper will describe what information can be captured, use that information to characterize some simple loads, and describe how that same information may be coordinated with other system measurements both to characterize new loads, and to more clearly identify scheduler shortcomings.