User:Bjacob/ArithmeticTimingDifferences: Difference between revisions

User:Bjacob/ArithmeticTimingDifferences (view source)

2 bytes added , 4 November 2013

Confirmed users

753

edits

@@ Line 52: / Line 52: @@
 The reason to compare these normal laws to the overall ''median'', not the overall ''mean'', is that the overall mean may fall in between the actual "normal" time range and some abnormal, much higher values obtained for certain operands. In that case, ''all'' actual timings would fall far away from the mean, which would cause all timings to be reported as abnormal! This problem is averted by using the median instead, which is by definition an actual timing, and is most stable in the face of outliers.
-Times are measured by clock_gettime using CLOCK_THREAD_CPUTIME_ID. In other words, this is per-thread CPU time. This should be as resilient to perturbations as it gets, especially as the core benchmarking loop is free of I/O or syscalls (no malloc) and relies on only at most two cache lines (it only accesses a few bytes on the stack, and in the worst case they would extend over one cache line boundary).
+Times are measured by clock_gettime using CLOCK_THREAD_CPUTIME_ID. In other words, this is per-thread CPU time. This should be as resilient to perturbations as it gets, especially as the core benchmarking loop is free of I/O or syscalls (no malloc) and relies on only at most two cache lines (it only accesses a few bytes on the stack, and in the worst case they would extend across one cache line boundary).
 == Caveats ==