Locality-Aware Task Scheduling and Data Distribution for OpenMP Programs on NUMA Systems and Manycore Processors

Figure 5

Latencies measured while accessing a cache line from different home cache access from tile 0 under isolation. Latencies increase in the presence of multiprogrammed workloads, OS, and hypervisor interference. Tile 63 runs dedicated system software and is excluded from measurement. Off-tile access takes 4–6 times longer.