readahead: make context readahead more conservative
authorFengguang Wu <fengguang.wu@intel.com>
Wed, 11 Sep 2013 21:21:47 +0000 (14:21 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 11 Sep 2013 22:57:39 +0000 (15:57 -0700)
commit2cad401801978b16ac6e43f10b8d60039670fcbc
tree1b3691a0b7660261ee2cacc7a7d9a76dbb9e0157
parent139c2d75b4e81d449d97f1f8188b84529eb56708
readahead: make context readahead more conservative

This helps performance on moderately dense random reads on SSD.

Transaction-Per-Second numbers provided by Taobao:

QPS case
-------------------------------------------------------
7536 disable context readahead totally
w/ patch: 7129 slower size rampup and start RA on the 3rd read
6717 slower size rampup
w/o patch: 5581 unmodified context readahead

Before, readahead will be started whenever reading page N+1 when it happen
to read N recently.  After patch, we'll only start readahead when *three*
random reads happen to access pages N, N+1, N+2.  The probability of this
happening is extremely low for pure random reads, unless they are very
dense, which actually deserves some readahead.

Also start with a smaller readahead window.  The impact to interleaved
sequential reads should be small, because for a long run stream, the the
small readahead window rampup phase is negletable.

The context readahead actually benefits clustered random reads on HDD
whose seek cost is pretty high.  However as SSD is increasingly used for
random read workloads it's better for the context readahead to concentrate
on interleaved sequential reads.

Another SSD rand read test from Miao

        # file size:        2GB
        # read IO amount: 625MB
        sysbench --test=fileio          \
                --max-requests=10000    \
                --num-threads=1         \
                --file-num=1            \
                --file-block-size=64K   \
                --file-test-mode=rndrd  \
                --file-fsync-freq=0     \
                --file-fsync-end=off    run

shows the performance of btrfs grows up from 69MB/s to 121MB/s, ext4 from
104MB/s to 121MB/s.

Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Tested-by: Tao Ma <tm@tao.ma>
Tested-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/readahead.c