zfs benchmarks and 7 disk raidz oddity

Andrew Snow andrew at modulus.org
Sat Sep 20 09:31:17 UTC 2008


Hi all,

I am running 8-CURRENT with ZFS patches on a 3ghz Core2Duo with Intel 
P45 chipset.  I took some benchmarks of ZFS on six old SATA disks and 
one PATA, onboard controllers only, and I expect you are interested to 
see the results.  The disks range in size from 200gb to 320gb.

I tried several configurations with quick and dirty testing. Listed 
below is the sequential megabytes/sec rating as measured by dd bs=1m for 
a 10GB file.  All ZFS settings were left at their defaults.


Conf                    Write   Read (MB/s)
------------------------------------
7 disk RAIDZ2           220     305
7 disk RAIDZ1            84     361
7 disk striped          318     409
7 disk stripe copies=2  140     164

6 disk RAIDZ2           173     260
6 disk RAIDZ1           238     307
6 disk striped          280     346
6 disk 2xRAIDZ1 striped 188     251
6 disk striped mirrors  164     323
6 disk stripe copies=2  151     179


A few notes:

1. using copies=2 is a nice way to be able to get RAID1-like mirroring 
reliability but on an odd number of disks.  However you take a 
noticeable performance penalty: Write speed is fine but read speed is 
almost half of what RAID0+1 achieved.

2. RAIDZ1 and RAIDZ2 are fast and efficient.  However in total it caused 
the system to use almost all of one CPU core during writing.

3. There seemed to be a bug with 7 disks and RAIDZ1 - the write 
performance was terrible!  When I ran "gstat" I noticed it was spending 
much time writing to only two disks most of the time, which became a 
serious bottleneck - the worst write score of the lot.  Read was fine. 
Perhaps the algorithm isn't optimised for choosing parity locations out 
of exactly 7 disks?



- Andrew



More information about the freebsd-fs mailing list