zfs discuss - Feb 2008 - ZFS raidz small IO write performance compared to raid controller

Hi all,

Does anyone have any data to show how ZFS raidz with the on-disk cache 
enabled for small, random IOs compares to a raid controller card with 
cache in raid 5.

I''m working on a very competitive RFP, and one thing that could give us
an advantage is the ability to remove this controller card.  I''ve never
measured this or seen it measured-- any pointers would be useful.  I 
believe the IOs are 8KB, the application is MySQL.

Thanks in advance,

- Matt

-- 
Matt Ingenthron - Web Infrastructure Solutions Architect
Sun Microsystems, Inc. - Global Systems Practice
http://blogs.sun.com/mingenthron/
email: matt.ingenthron at sun.com             Phone: 310-242-6439

For small random I/O operations I would expect a substantial performance penalty
for ZFS. The reason is that RAID-Z is more akin to RAID-3 than RAID-5; each read
and write operation touches all of the drives. RAID-5 allows multiple I/O
operations to proceed in parallel since each read and write operation touches
only 2 drives.

As always, benchmarking the application is best.  :-)
 
 
This message posted from opensolaris.org

Matt Ingenthron wrote:
> Hi all,
>
> Does anyone have any data to show how ZFS raidz with the on-disk cache 
> enabled for small, random IOs compares to a raid controller card with 
> cache in raid 5.
>
> I''m working on a very competitive RFP, and one thing that could
give us
> an advantage is the ability to remove this controller card.  I''ve
never
> measured this or seen it measured-- any pointers would be useful.  I 
> believe the IOs are 8KB, the application is MySQL.
>   
In general, low cost and fast are mutually exclusive.  To get lots of 
database
performance you tend to need lots of disks.  Cache effects are secondary.
Also, if you need performance, RAID-1 beats RAID-5.

Anton B. Rang wrote:
> For small random I/O operations I would expect a substantial performance
penalty for ZFS. The reason is that RAID-Z is more akin to RAID-3 than RAID-5;
each read and write operation touches all of the drives. RAID-5 allows multiple
I/O operations to proceed in parallel since each read and write operation
touches only 2 drives.
>   
There are a lot of caveats glossed over here.  The main RAID-5 penalty
for writes is the read-modify write sequence which is often required,
but usually nicely hidden by a RAID controller with nonvolatile cache.
For raidz, it is a little more complex because it is possible that a write
will only cause 2 iops or a single write (2+ physical writes) may contain
many database blocks written sequentially.  The net effect of these
complexities is that write performance is very, very difficult to predict.
Read performance is more consistently predictable, but only for the
case where reads are aligned and caches are always missed (which suck
anyway).
> As always, benchmarking the application is best.  :-)
>  
Absolutely!
 -- richard