系统技术非业余研究

Linux的高速缓存pagecache对性能的影响至关重要，但是实际系统中我们的利用率如何呢，特别是具体到每个设备的利用情况。

从下图我们可以很清楚的看到：

我们知道IO请求由vfs发起，经过pagecache缓存，挡不住的就落实到io设备去，那么统计这个利用率就很简单。 我们只要知道挡不住的IO的比例就好了。

我写了个systemtap脚本来解决这个问题：

$ uname -r
2.6.18-164.el5
 
$ sudo rpm -i kernel-debuginfo-common-2.6.18-164.el5.x86_64.rpm
$ sudo rpm -i kernel-debuginfo-2.6.18-164.el5.x86_64.rpm  

$ cat >pcmr.stp
global io_stat, blk_stat

global RD_CNT=2, WR_CNT=3

probe vfs.read.return {
  if ($return>0 && devname!="N/A") {
      io_stat[devname, 0] += $return
      io_stat[devname, RD_CNT] ++;
  }
}

probe ioblock.request {
  blk_stat[devname, rw] += size
  blk_stat[devname, rw+RD_CNT] ++;
} 

probe vfs.write.return {
  if ($return>0 && devname!="N/A") {
      io_stat[devname, 1] += $return
      io_stat[devname, WR_CNT] ++;
  }
}


global _io,_blk

probe timer.ms(5000) { 

foreach ([devname, action] in io_stat)
{
        _io[action] += io_stat[devname, action];
        _blk[action] += blk_stat[devname, action]
}

  if (_io[0] + _io[1]) {

    printf("\n%-25s, %s%db/%d, %s%db/%d, %s%db/%d, %s%db/%d, %s%d%%\n\n",
           ctime(gettimeofday_s()),
           "IO Read:", _io[0], _io[RD_CNT],
           "IO Write:", _io[1],_io[WR_CNT], 
           "Blk Read:", _blk[0], _blk[RD_CNT], 
           "Blk Write:", _blk[1],_blk[WR_CNT], 
           "PCMR:", (_blk[0]+_blk[1]) * 100 / (_io[0] + _io[1]))

        delete _io
        delete _blk
    
    /* print header */
    printf("%8s %2s %19s %19s %5s\n",
           "DEVICE","T", "IO", "BLK", "PCMR")
  }

  /* print top ten I/O */
  foreach ([devname, action] in io_stat- limit 10)
     if(action < RD_CNT) 
        printf("%8s %2s %11db/%6d %11db/%6d %4d%%\n",
           devname, action?"W":"R",
           io_stat[devname, action],io_stat[devname, action+RD_CNT],
           blk_stat[devname, action],blk_stat[devname, action+RD_CNT],
           (blk_stat[devname, action] * 100) / io_stat[devname, action]
           )

  /* clear data */
  delete io_stat
  delete blk_stat
}

probe begin
{
println("::");
}

CTRL+D

我们来演示下：

$ sudo stap pcmr.stp    
::
...
Wed Jun  1 09:16:16 2011 , IO Read:551320328b/124024, IO Write:0b/0, Blk Read:103219200b/17536, Blk Write:0b/0, PCMR:18%

  DEVICE  T                  IO                 BLK  PCMR
    fioa  R   493284566b/119460   103219200b/ 17536   20%
    sda2  R    41100816b/   624           0b/     0    0%
    sda5  R    16934946b/  3940           0b/     0    0%

Wed Jun  1 09:16:21 2011 , IO Read:554808170b/123568, IO Write:2295b/9, Blk Read:101883904b/17296, Blk Write:135168b/33, PCMR:18%

  DEVICE  T                  IO                 BLK  PCMR
    fioa  R   491003095b/118908   101543936b/ 17259   20%
    sda3  R    42119596b/   136           0b/     0    0%
    sda5  R    17730097b/  4061      163840b/    18    0%
    sda1  R     2399374b/    33           0b/     0    0%
    sda2  R     1543938b/   424      176128b/    19   11%
    sda6  R       12070b/     6           0b/     0    0%
    sda3  W        2295b/     9      135168b/    33 5889%
...

参数我们来解释下：
IO Read: vfs层面看到的IO读/次数
IO Write: vfs层面看到的IO写/次数
Blk Read: 块层面看到的IO读/次数
Blk Write: 块层面看到的IO写/次数
PCMR: pagecache miss rate =100 * blk / io, 由于文件读写是以page为单位，所以可能>100%
第一行是汇总统计，每隔5秒打印一次。
总结： PCMR越小越好，说明高速缓存利用率高。

玩得开心！