)]}'
{
  "commit": "ea3f01f8afd3bc5daff915cc4ea5cc5ea9e7d427",
  "tree": "e2330451fc9d6cc02840c31be3e4956db7974a76",
  "parents": [
    "b3137bc8e77962a8e3b4dfdc1bcfd38e437bd278"
  ],
  "author": {
    "name": "Ingo Molnar",
    "email": "mingo@elte.hu",
    "time": "Thu May 29 14:32:23 2008 +0200"
  },
  "committer": {
    "name": "Ingo Molnar",
    "email": "mingo@elte.hu",
    "time": "Thu May 29 14:46:30 2008 +0200"
  },
  "message": "sched: re-tune NUMA topologies\n\nimprove the sysbench ramp-up phase and its peak throughput on\na 16way NUMA box, by turning on WAKE_AFFINE:\n\n             tip/sched   tip/sched+wake-affine\n-------------------------------------------------\n    1:             700              830    +15.65%\n    2:            1465             1391    -5.28%\n    4:            3017             3105    +2.81%\n    8:            5100             6021    +15.30%\n   16:           10725            10745    +0.19%\n   32:           10135            10150    +0.16%\n   64:            9338             9240    -1.06%\n  128:            8599             8252    -4.21%\n  256:            8475             8144    -4.07%\n-------------------------------------------------\n  SUM:           57558            57882    +0.56%\n\nthis change also improves lat_ctx from 6.69 usecs to 1.11 usec:\n\n  $ ./lat_ctx -s 0 2\n  \"size\u003d0k ovr\u003d1.19\n  2 1.11\n\n  $ ./lat_ctx -s 0 2\n  \"size\u003d0k ovr\u003d1.22\n  2 6.69\n\nin sysbench it\u0027s an overall win with some weakness at the lots-of-clients\nside. That happens because we now under-balance this workload\na bit. To counter that effect, turn on NEWIDLE:\n\n              wake-idle          wake-idle+newidle\n -------------------------------------------------\n     1:             830              834    +0.43%\n     2:            1391             1401    +0.65%\n     4:            3105             3091    -0.43%\n     8:            6021             6046    +0.42%\n    16:           10745            10736    -0.08%\n    32:           10150            10206    +0.55%\n    64:            9240             9533    +3.08%\n   128:            8252             8355    +1.24%\n   256:            8144             8384    +2.87%\n -------------------------------------------------\n   SUM:           57882            58591    +1.21%\n\nas a bonus this not only improves the many-clients case but\nalso improves the (more important) rampup phase.\n\nsysbench is a workload that quickly breaks down if the\nscheduler over-balances, so since it showed an improvement\nunder NEWIDLE this change is definitely good.\n",
  "tree_diff": [
    {
      "type": "modify",
      "old_id": "4bb7074a2c3a5d018ff13eee399d82a2be6c10e6",
      "old_mode": 33188,
      "old_path": "include/linux/topology.h",
      "new_id": "24f3d2282e1184c3741e7d99f212ca4079f3e723",
      "new_mode": 33188,
      "new_path": "include/linux/topology.h"
    }
  ]
}