Change Details

The current vmmeter struct is subject to extreme false sharing due to several atomic counters sharing cachelines. This severely reduces performance on fork/exec heavy workloads on high core boxes. pig1 (westmere 4 * 10 * 2) doing -j 80 buildkernel on tmpfs goes from: 3599.82s user 1313.64s system 5485% cpu 1:29.58 total to: 3613.17s user 1124.48s system 5500% cpu 1:26.13 total i.e. 3 seconds less real time and ~1300 -> ~1100 system time. The patch goes for the most trivial approach of moving relevant counters with __exclusive_cache_line.

The current vmmeter struct is subject to extreme false sharing due to several atomic counters sharing cachelines. This severely reduces performance on fork/exec heavy workloads on high core boxes. pig1 (westmere 4 * 10 * 2) doing -j 80 buildkernel on tmpfs goes from: 3599.82s user 1313.64s system 5485% cpu 1:29.58 total to: 3613.17s user 1124.48s system 5500% cpu 1:26.13 total i.e. 3 seconds less real time and ~1300 -> ~1100 system time. The patch goes for the most trivial approach of moving relevant counters out and annotating them with __exclusive_cache_line. I'm not attached to the new name prefix (vm_) nor names of padding fields. I just would like to see the split (or an equivalent) in.

The current vmmeter struct is subject to extreme false sharing due to several atomic counters sharing cachelines. This severely reduces performance on fork/exec heavy workloads on high core boxes. pig1 (westmere 4 * 10 * 2) doing -j 80 buildkernel on tmpfs goes from: 3599.82s user 1313.64s system 5485% cpu 1:29.58 total to: 3613.17s user 1124.48s system 5500% cpu 1:26.13 total i.e. 3 seconds less real time and ~1300 -> ~1100 system time. The patch goes for the most trivial approach of moving relevant counters with __exclusive_cache_lineout and annotating them with __exclusive_cache_line. I'm not attached to the new name prefix (vm_) nor names of padding fields. I just would like to see the split (or an equivalent) in.