[zfs-discuss] Slow read performance
alex.vodeyko at gmail.com
Mon Apr 2 09:03:00 EDT 2018
Yes sure, "dd" is just a quick check.
Because if I get with "dd" 1+ GB/s reads from 10 drives and < 0.5 GB/s
from 30 (and even 60) drives - then something is fundamentally wrong.
2018-04-02 15:37 GMT+03:00 Gordan Bobic <gordan.bobic at gmail.com>:
> You do realize that measuring sequential I/O performance with dd will
> produce readings that are in no meaningful way correlatable to any
> multi-user workload you are likely to throw at it, especially something as
> metadata intensive (small random I/O) as Lustre, right?
> On Mon, 2 Apr 2018, 12:05 Alex Vodeyko via zfs-discuss,
> <zfs-discuss at list.zfsonlinux.org> wrote:
>> Yes, I'm now running zpool with two 12+3 raidz3 VDEVs. Read
>> performance is definitely better than zpool with three 8+2 raidz2
>> VDEVs (And still major read performance improvement comes from ashift
>> = 9):
>> - "dd" ashift = 9: write = 2.3 GB/s, read = 1.8 GB/s
>> - "dd" ashift = 12: write = 2.8 GB/s, read = 1.2 GB/s
>> I decided to have 2x 30-drive zpools (because it seems not too much
>> performance difference between 30 and 60 drives in all my tested zpool
>> configs). I will use Lustre filesystem, so decided to have 2x Lustre
>> OSS servers with 30 drives each.
>> Still choosing the correct zpool layout - from the benchmarks it seems
>> raidz3 performance is only 100 MB/s worse.
>> But it would be great to get the advice on the best zpool layout
>> (raidz2 vs raidz3, 12+2, 13+2, 12+3) for 30 drives.
>> I'm now for two 12+3 raidz3 VDEVs with ashift=9.
>> 2018-04-01 23:20 GMT+03:00 Andreas Dilger <adilger at dilger.ca>:
>> > On Mar 31, 2018, at 11:23 PM, Alex Vodeyko <alex.vodeyko at gmail.com>
>> > wrote:
>> >> To remind - all of the above came from zpool of six 8+2 raidz2 (60
>> >> drives total)
>> >> For comparison I've created one zpool with single 8+2 raidz2 (10
>> >> drives) and rerun tests on it, so:
>> > Have you tried a different geometry, like 5x 10+2 RAID-Z2 VDEVs? At one
>> > time there was a bug in the block allocation code that made 8+2 not work as
>> > well as 9+2 or 7+2. That _should_ have been fixed in the 0.7.x version you
>> > are running, but there might still be some problems.
>> > Cheers, Andreas
>> zfs-discuss mailing list
>> zfs-discuss at list.zfsonlinux.org
More information about the zfs-discuss