The reason I would like the "limited range" averages is to rule out excessive peaks. When timing small and very intensive routines, a task switch IS important, because it flushes the cache, and make filter perform considerably slower.
Of course cache flushes occur in real life, but these cases "doesn't count" when optimizing inner loops. However if the top 10% is consistently slower than the center 50% execution time without any obvious reason it might be worth a look.
Furthermore nonlinear access and first frames are often slower - so if they were in the top 25% - these cases would also be taken out of the equation.
The reason I would avoid the lower 25% is for cache hits not to count in the average.
__________________
Regards, sh0dan // VoxPod
|