Just select a random offset in order to mitigate the unpredictability
of cache associativity effects when dealing with different physical
memory fragmentation (for PIPT caches). We are reporting the "best"
measured latency, some offsets may be better than the others.