We also compared the number of cache misses expected to be caused by our benchmark to those actually caused, using hardware performance monitoring counters. We found that for a system with the LRU replacement policy, our estimate is highly accurate while it fails for a system with a non-LRU policy.