深入理解 linux磁盘顺序写、随机写

一、前言

● 随机写会导致磁头不停地换道，造成效率的极大降低；顺序写磁头几乎不用换道，或者换道的时间很短
● 本文来讨论一下两者具体的差别以及相应的内核调用

二、环境准备

组件	版本
OS	Ubuntu 16.04.4 LTS
fio	2.2.10

三、fio介绍

通过fio测试，能够反映在读写中的状态，我们需要重点关注fio的输出报告中的几个关键指标：
slat ：是指从 I/O 提交到实际执行 I/O 的时长（Submission latency）
clat ：是指从 I/O 提交到 I/O 完成的时长（Completion latency）
lat ：指的是从 fio 创建 I/O 到 I/O 完成的总时长
bw ：吞吐量
iops ：每秒 I/O 的次数

四、同步写测试

（1）同步随机写

主要采用fio作为测试工具，为了能够看到系统调用，使用strace工具，命令看起来是这样：

先来测试一个随机写

strace -f -tt -o /tmp/randwrite.log -D fio -name=randwrite -rw=randwrite 
-direct=1 -bs=4k -size=1G -numjobs=1  -group_reporting -filename=/tmp/test.db

提取关键信息

root@wilson-ubuntu:~# strace -f -tt -o /tmp/randwrite.log -D fio -name=randwrite -rw=randwrite 
> -direct=1 -bs=4k -size=1G -numjobs=1  -group_reporting -filename=/tmp/test.db
randwrite: (g=0): rw=randwrite, bs=4K-4K/4K-4K/4K-4K, ioengine=sync, iodepth=1
fio-2.2.10
Starting 1 process
...
randwrite: (groupid=0, jobs=1): err= 0: pid=26882: Wed Aug 14 10:39:02 2019
  write: io=1024.0MB, bw=52526KB/s, iops=13131, runt= 19963msec
    clat (usec): min=42, max=18620, avg=56.15, stdev=164.79
     lat (usec): min=42, max=18620, avg=56.39, stdev=164.79
...
    bw (KB  /s): min=50648, max=55208, per=99.96%, avg=52506.03, stdev=1055.83
...

Run status group 0 (all jobs):
  WRITE: io=1024.0MB, aggrb=52525KB/s, minb=52525KB/s, maxb=52525KB/s, mint=19963msec, maxt=19963msec

Disk stats (read/write):
...
  sda: ios=0/262177, merge=0/25, ticks=0/7500, in_queue=7476, util=36.05%

列出了我们需要重点关注的信息：
（1）clat ，平均时长56ms左右
（2）lat ，平均时长56ms左右
（3）bw ，吞吐量，大概在52M左右

再来看内核调用信息：

root@wilson-ubuntu:~# more /tmp/randwrite.log
...
26882 10:38:41.919904 lseek(3, 665198592, SEEK_SET) = 665198592
26882 10:38:41.919920 write(3, "220240@6371341277>2003631202273203436H26P34027737033030e30"..., 4096) = 4096
26882 10:38:41.919969 lseek(3, 4313088, SEEK_SET) = 4313088
26882 10:38:41.919985 write(3, "220240@6371341277>2003631202273203436H26P34027737033030e30"..., 4096) = 4096
26882 10:38:41.920032 lseek(3, 455880704, SEEK_SET) = 455880704
26882 10:38:41.920048 write(3, "220240@6371341277>2003631202273203436H26P34027737033030e30"..., 4096) = 4096
26882 10:38:41.920096 lseek(3, 338862080, SEEK_SET) = 338862080
26882 10:38:41.920112 write(3, "220240@6371341277>240224202273203436H26P34027737033030e30"..., 4096) = 4096
26882 10:38:41.920161 lseek(3, 739086336, SEEK_SET) = 739086336
26882 10:38:41.920177 write(3, "220240@6371341277>240224202273203436H26P34027737033030e30"..., 4096) = 4096
26882 10:38:41.920229 lseek(3, 848175104, SEEK_SET) = 848175104
26882 10:38:41.920245 write(3, "220240@6371341277>240224202273203436H26P34027737033030e30"..., 4096) = 4096
26882 10:38:41.920296 lseek(3, 1060147200, SEEK_SET) = 1060147200
26882 10:38:41.920312 write(3, "220240@6371341277>240224202273203436H26P34027737033030e30"..., 4096) = 4096
26882 10:38:41.920362 lseek(3, 863690752, SEEK_SET) = 863690752
26882 10:38:41.920377 write(3, "220240@6371341277>240224202273203436H26P34027737033030e30"..., 4096) = 4096
26882 10:38:41.920428 lseek(3, 279457792, SEEK_SET) = 279457792
26882 10:38:41.920444 write(3, "220240@6371341277>240224202273203436H26P34027737033030e30"..., 4096) = 4096
26882 10:38:41.920492 lseek(3, 271794176, SEEK_SET) = 271794176
26882 10:38:41.920508 write(3, "220240@6371341277>240224202273203436H26P34027737033030e30"..., 4096) = 4096
26882 10:38:41.920558 lseek(3, 1067864064, SEEK_SET) = 1067864064
26882 10:38:41.920573 write(3, "220240@6371341277>240224202273203436H26P34027737033030e30"..., 4096) = 4096
...

随机读每一次写入之前都要通过lseek去定位当前的文件偏移量

同步顺序写

用刚才的方法来测试顺序写

root@wilson-ubuntu:~# strace -f -tt -o /tmp/write.log -D fio -name=write -rw=write 
-direct=1 -bs=4k -size=1G -numjobs=1  -group_reporting -filename=/tmp/test.db
write: (g=0): rw=write, bs=4K-4K/4K-4K/4K-4K, ioengine=sync, iodepth=1
fio-2.2.10
Starting 1 process
Jobs: 1 (f=1): [W(1)] [100.0% done] [0KB/70432KB/0KB /s] [0/17.7K/0 iops] [eta 00m:00s]
write: (groupid=0, jobs=1): err= 0: pid=27005: Wed Aug 14 10:53:02 2019
  write: io=1024.0MB, bw=70238KB/s, iops=17559, runt= 14929msec
    clat (usec): min=43, max=7464, avg=55.95, stdev=56.24
     lat (usec): min=43, max=7465, avg=56.15, stdev=56.25
...
    bw (KB  /s): min=67304, max=72008, per=99.98%, avg=70225.38, stdev=1266.88
...

Run status group 0 (all jobs):
  WRITE: io=1024.0MB, aggrb=70237KB/s, minb=70237KB/s, maxb=70237KB/s, mint=14929msec, maxt=14929msec

Disk stats (read/write):
...
  sda: ios=0/262162, merge=0/10, ticks=0/6948, in_queue=6932, util=46.49%

可以看到：
吞吐量提升至70M左右

再来看一下内核调用：

root@wilson-ubuntu:~# more /tmp/write.log
...
27046 10:54:28.194508 write(3, "2330016360	32017"..., 4096) = 4096
27046 10:54:28.194568 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.194627 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.194687 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.194747 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.194807 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.194868 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.194928 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.194988 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.195049 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.195110 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.195197 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.195262 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.195330 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.195426 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.195497 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.195567 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.195637 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.195704 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.195757 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.195807 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.195859 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.195910 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.195961 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.196012 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.196062 write(3, "23300162202432017"..., 4096) = 4096
27046 10:54:28.196112 write(3, "2330016 2632017"..., 4096) = 4096
27046 10:54:28.196162 write(3, "2330016 2632017"..., 4096) = 4096
27046 10:54:28.196213 write(3, "2330016 2632017"..., 4096) = 4096
27046 10:54:28.196265 write(3, "2330016 2632017"..., 4096) = 4096
27046 10:54:28.196314 write(3, "2330016 2632017"..., 4096) = 4096
27046 10:54:28.196363 write(3, "2330016 2632017"..., 4096) = 4096
27046 10:54:28.196414 write(3, "2330016 2632017"..., 4096) = 4096
27046 10:54:28.196472 write(3, "2330016 2632017"..., 4096) = 4096
27046 10:54:28.196524 write(3, "2330016 2632017"..., 4096) = 4096
27046 10:54:28.196573 write(3, "2330016 2632017"..., 4096) = 4096
...

由于顺序读，不需要反复定位文件偏移量，所以能够专注于写操作

五、slat指标

从上面的测试，我们在fio的测试报告中，并没有发现slat的身影，那是由于上述都是同步操作，对同步 I/O 来说，由于 I/O 提交和 I/O 完成是一个动作，所以 slat 实际上就是 I/O 完成的时间

异步顺序写，将同步顺序写的命令添加-ioengine=libaio：

root@wilson-ubuntu:~# fio -name=write -rw=write -ioengine=libaio -direct=1 -bs=4k -size=1G -numjobs=1  -group_reporting -filename=/tmp/test.db
write: (g=0): rw=write, bs=4K-4K/4K-4K/4K-4K, ioengine=libaio, iodepth=1
fio-2.2.10
Starting 1 process
Jobs: 1 (f=1): [W(1)] [100.0% done] [0KB/119.3MB/0KB /s] [0/30.6K/0 iops] [eta 00m:00s]
write: (groupid=0, jobs=1): err= 0: pid=27258: Wed Aug 14 11:14:36 2019
  write: io=1024.0MB, bw=120443KB/s, iops=30110, runt=  8706msec
    slat (usec): min=3, max=70, avg= 4.31, stdev= 1.56
    clat (usec): min=0, max=8967, avg=28.13, stdev=55.68
     lat (usec): min=22, max=8976, avg=32.53, stdev=55.72
...
    bw (KB  /s): min=118480, max=122880, per=100.00%, avg=120467.29, stdev=1525.68
...

Run status group 0 (all jobs):
  WRITE: io=1024.0MB, aggrb=120442KB/s, minb=120442KB/s, maxb=120442KB/s, mint=8706msec, maxt=8706msec

Disk stats (read/write):
...
  sda: ios=0/262147, merge=0/1, ticks=0/6576, in_queue=6568, util=74.32%

可以看到，slat指标出现，lat 近似等于 slat + clat 之和（avg平均值）；并且换成异步io之后，吞吐量得到了极大的提升，120M左右

六、总结

● fio应该作为磁盘的baseline工具，拿到机器（物理机或者云机器）都应该第一时间对机器的磁盘做一个基线测试，做到心中有数
● 本文所有的测试，都是绕开了缓存，在实际应用中需要将缓存的影响考虑进去

---

至此，本文结束
在下才疏学浅，有撒汤漏水的，请各位不吝赐教...