Interestingly, a soft page fault can be caused not only by a real page fault event, but also by NUMA autobalancing. Which actually could be a major source of ke

Interestingly, a soft page fault can be caused not only by a real page fault event, but also by NUMA autobalancing. Which actually could be a major source of kernel interference to purely user space processes. I sometimes see these NUMA-originated soft page faults very high up in a (CPU/cache) profiler for sensitive real-time processes that should avoid contexts switches and L1/L2 cache pollution at all costs. As a side note NUMA autobalancing can be disabled by running: echo 0 > /proc/sys/kernel /numa_balancing_scan_period_min_ms echo 0 > /proc/sys/kernel/numa_balancing_scan_period_max_ms echo 0 > /proc/sys/kernel/numa_balancing_scan_size_mb echo 1000000 > /proc/sys/kernel/numa_balancing_scan_period_min_ms Or booting the box with the kernel command line that includes: numa_balancing=disable