利用Trim/discard挂载rbd块设备

在http://cephnotes.ksperis.com/blog/2014/12/18/use-discard-with-krbd-client-since-kernel-3-dot-18/有详细的描述，试验了一下，对于小文件的删除确实有帮助，记录如下。

1. 如果不利用discard选项挂载rbd块设备，会有如下问题

$ rbd create rbd/myrbd --size=20480
$ mkfs.xfs /dev/rbd0
$ rbd diff rbd/myrbd | awk '{ SUM += $2 } END { print SUM/1024/1024 " MB" }'
14.4062 MB

$ mount /dev/rbd0 /mnt/myrbd
$ dd if=/dev/zero of=/mnt/myrbd/testfile bs=1M count=1024
$ rbd diff rbd/myrbd | awk '{ SUM += $2 } END { print SUM/1024/1024 " MB" }'
1038.41 MB

$ rm /mnt/myrbd/testfile
$ rbd diff rbd/myrbd | awk '{ SUM += $2 } END { print SUM/1024/1024 " MB" }'
1038.41 MB

你会发现，rbd pool并没有变小。之前好像做过类似的测试，需要重启一下ceph或系统才能反映出改变，记不清了。

2. 利用discard选项挂载rbd块设备

mount -o discard /dev/rbd0 /mnt/myrbd
重复上述操作，再执行以下操作：
$ fstrim /mnt/myrbd
$ rbd diff rbd/myrbd | awk '{ SUM += $2 } END { print SUM/1024/1024 " MB" }'
10.6406 MB

这样就恢复正常了。

两种不同挂载方式的性能测试如下：

1. 未使用discard选项

$ mount /dev/rbd0 /mnt/rbd0

$ mkdir testdir; cd testdir
$ dd if=/dev/zero of=mainfile bs=1M count=200
$ split -b 4048 -a 7 mainfile; sync               # 4k file / ~51k files
$ cd ..
$ time rm -rf testdir; time sync

[root@mon0 nfs]# time rm -rf testdir/; time sync

real	0m21.866s
user	0m0.155s
sys	0m5.466s

real	0m0.897s
user	0m0.001s
sys	0m0.011s

2. 使用discard选项

只给出时间
[root@mon0 nfs]# time rm -rf testdir; time sync

real	0m8.791s
user	0m0.100s
sys	0m2.858s

real	0m2.367s
user	0m0.001s
sys	0m0.001s

性能有明显提升，作者的原话是：

In the case of intensive use of the file system, with many small file, it may be more advantageous to use fstrim, for example once a day.