drive performance

[Eric Shubert]

der.hans wrote:
> moin moin,
> 
> We're using SATA II ( 3.0 GB/s ) drives. Some of our servers consistently
> have iowait issues. Those machines also can't tolerate drive search tools
> such as find and du.
> 
> Looking at one that was having problems today. top is showing wait
> constantly over 5% and up to 15%. It was stuck over 10% earlier today and
> up to 50%. It's a dual-proc quad-core system.
> 
> iowait is showing write staying below 3 MB/s and currently mostly less
> than 1 MB/s. Reads are almost non-existent.
> 
> Avg wait is jumping up to 800 ms, sometimes higher.
> 
> This machine is mostly only writing out MySQL changes and logs.
> 
> Should writing 3GB/s really cause a lot of io wait?
> 
> Hmm, I see the filesystem is not mounted noatime, changing that now.
> Didn't immediately change io wait.
> 
> A java process is eating most of the 64GB of RAM, but the swap partition
> is not showing any disk activity. The java process is tied to another disk
> as well.
> 
> This machine does not have an SAS controller, but I am considering moving
> to SAS for future machines, budget willing. We might just move to RAID 0
> for the DB. The java proc is already writing to RAID 0 in most cases.
> 
> I'm going to run some iozone or bonnie++ checks on a similar machine that
> has not yet deployed.
> 
> $sudo hdparm /dev/sda
> 
> /dev/sda:
>  IO_support   =  0 (default 16-bit)
>  readonly     =  0 (off)
>  readahead    = 256 (on)
>  geometry     = 19457/255/63, sectors = 312581808, start = 0
> 
> ciao,
> 
> der.hans

Which elevator is being used?
(cat /sys/block/sda/queue/scheduler)

Using async i/o? Synchronous writes are very slow.

I presume ext3, since you referred to noatime. If you have directories 
with many files (e.g. Maildir), dir_index option can help.

-- 
-Eric 'shubes'