xfs
[Top] [All Lists]

Re: Performance problem - reads slower than writes

To: Brian Candler <B.Candler@xxxxxxxxx>
Subject: Re: Performance problem - reads slower than writes
From: Dave Chinner <david@xxxxxxxxxxxxx>
Date: Wed, 1 Feb 2012 07:25:26 +1100
Cc: xfs@xxxxxxxxxxx
In-reply-to: <20120131141604.GB46571@xxxxxxxx>
References: <20120130220019.GA45782@xxxxxxxx> <20120131020508.GF9090@dastard> <20120131103126.GA46170@xxxxxxxx> <20120131141604.GB46571@xxxxxxxx>
User-agent: Mutt/1.5.21 (2010-09-15)
On Tue, Jan 31, 2012 at 02:16:04PM +0000, Brian Candler wrote:
> Updates:
> 
> (1) The bug in bonnie++ is to do with memory allocation, and you can work
> around it by putting '-n' before '-s' on the command line and using the same
> custom chunk size before both (or by using '-n' with '-s 0')
> 
> # time bonnie++ -d /data/sdc -n 98:800k:500k:1000:32k -s 16384k:32k -u root
> 
> Version  1.96       ------Sequential Output------ --Sequential Input- 
> --Random-
> Concurrency   1     -Per Chr- --Block-- -Rewrite- -Per Chr- --Block-- 
> --Seeks--
> Machine   Size:chnk K/sec %CP K/sec %CP K/sec %CP K/sec %CP K/sec %CP  /sec 
> %CP
> storage1    16G:32k  2061  91 101801   3 49405   4  5054  97 126748   6 130.9 
>   3
> Latency             15446us     222ms     412ms   23149us   83913us     452ms
> Version  1.96       ------Sequential Create------ --------Random 
> Create--------
> storage1            -Create-- --Read--- -Delete-- -Create-- --Read--- 
> -Delete--
> files:max:min        /sec %CP  /sec %CP  /sec %CP  /sec %CP  /sec %CP  /sec 
> %CP
> 98:819200:512000/1000   128   3    37   1 10550  25   108   3    38   1  8290 
>  33
> Latency              6874ms   99117us   45394us    4462ms   12582ms    4027ms
> 1.96,1.96,storage1,1,1328002525,16G,32k,2061,91,101801,3,49405,4,5054,97,126748,6,130.9,3,98,819200,512000,,1000,128,3,37,1,10550,25,108,3,38,1,8290,33,15446us,222ms,412ms,23149us,83913us,452ms,6874ms,99117us,45394us,4462ms,12582ms,4027ms
> 
> This shows that using 32k transfers instead of 8k doesn't really help; I'm
> still only seeing 37-38 reads per second, either sequential or random.

Right, because it is doing buffered IO and reading and writing into
the page cache for small Io sizes is much faster than waiting for
physical IO. Hence there is much less of a penalty for small
buffered IOs compared.

> (2) In case extents aren't being kept in the inode, I decided to build a
> filesystem with '-i size=1024'
> 
> # time bonnie++ -d /data/sdb -n 98:800k:500k:1000:32k -s0 -u root
> 
> Version  1.96       ------Sequential Create------ --------Random 
> Create--------
> storage1            -Create-- --Read--- -Delete-- -Create-- --Read--- 
> -Delete--
> files:max:min        /sec %CP  /sec %CP  /sec %CP  /sec %CP  /sec %CP  /sec 
> %CP
> 98:819200:512000/1000   110   3   131   5  3410  10   110   3    33   1   387 
>   1
> Latency              6038ms   92092us   87730us    5202ms     117ms    7653ms
> 1.96,1.96,storage1,1,1328003901,,,,,,,,,,,,,,98,819200,512000,,1000,110,3,131,5,3410,10,110,3,33,1,387,1,,,,,,,6038ms,92092us,87730us,5202ms,117ms,7653ms
> 
> Wow! The sequential read just blows away the previous results. What's even
> more amazing is the number of transactions per second reported by iostat
> while bonnie++ was sequentially stat()ing and read()ing the files:

The only thing changing the inode size will have affected is the
directory structure - maybe your directories are small enough to fit
in line, or the inode is large enough to keep it in extent format
rather than a full btree. In either case, though, the directory
lookup will require less IO.

> 
> # iostat 5
> ...
> sdb             820.80     86558.40         0.00     432792          0
>                   !!
> 
> 820 tps on a bog-standard hard-drive is unbelievable, although the total
> throughput of 86MB/sec is.  It could be that either NCQ or drive read-ahead
> is scoring big-time here.

See my previous explanation of adjacent IOs not needing seeks. All
you've done is increase the amount of IO needed to read and write
inodes because the inode cluster size is a fixed 8k. That means you
now need to do 8 adjacent IOs to read a 64 inode chunk instead of 2
adjecent IOs when you have 256 byte inodes. And because they are
adjacent IOs, they will hit the drive cache and so not require
physical IO to be done. Hence you can get much "higher" Io
throughput without actually doing any more physical IO....

> However during random stat()+read() the performance drops:
> 
> # iostat 5
> ...
> sdb             225.40     21632.00         0.00     108160          0

Because it is now reading random inodes so not reading adjacent
8k inode clusters all the time.

> 
> Here we appear to be limited by real seeks. 225 seeks/sec is still very good

That number indicates 225 IOs/s, not 225 seeks/s.

> for a hard drive, but it means the filesystem is generating about 7 seeks
> for every file (stat+open+read+close).  Indeed the random read performance

7 IOs for every file.

> appears to be a bit worse than the default (-i size=256) filesystem, where
> I was getting 25MB/sec on iostat, and 38 files per second instead of 33.

Right, because it is taking more seeks to read the inodes because they
are physically further apart.

> There are only 1000 directories in this test, and I would expect those to
> become cached quickly.

Doubtful. There's plenty of page cache pressure (500-800k) per inode
read (maybe 16k of cached metadata all up) so there's enough memory
pressure to prevent the directory structure from staying memory
resident.

> It looks like I need to get familiar with xfs_db and
> http://oss.sgi.com/projects/xfs/papers/xfs_filesystem_structure.pdf
> to find out what's going on.

It's pretty obvious to me what is happening. :/ I think that you first
need to understand exactly what the tools you are already using are
actually telling you, then go from there...

Cheers,

Dave.
-- 
Dave Chinner
david@xxxxxxxxxxxxx

<Prev in Thread] Current Thread [Next in Thread>