On 01/20/2018 10:20 AM, Chris Ferebee wrote:
Am 19.01.2018 um 19:25 schrieb Ian Collins <[email protected]>:
There never has been any need to mirror L2ARC device, you might get a small
performance gain reading from a mirror. Now logs are recoverable, there is
little benefit from mirroring log devices.
Ian,
Can you comment on
<https://docs.joyent.com/private-cloud/troubleshooting/disk-replacement>
which reads like recent, official guidance from Joyent and argues for mirrored
slog devices? (My emphasis below.) Just trying to get a handle on best
practice. If in fact SmartOS does handle a slog failure gracefully—happily,
I’ve never had one—your arguments certainly make sense.
Best,
Chris
+++ quote: +++
Working with ZIL logs
ZIL log devices are a special case in ZFS. They 'front' synchronous writes to
the pool: slower sync writes get pushed to the pool and are effectively cached
to fast temporary storage to allow storage consumers to continue, with the
mechanisms for the ZIL flushing transactions from the log to permanent storage
in bursts.
This in effect makes the ZIL a dangerous single point of failure to the pool in
certain situations. For instance, if a single log device fails, it effectively
cuts out the middle of the data pipeline while losing any transactions
in-flight: data which was already written to the pool and present on the log,
but not yet committed to persistent storage, will be lost.
RUNNING MIRRORED ZIL LOG DEVICES IS HIGHLY RECOMMENDED and mitigates this
single point of failure. Working with ZIL log mirrors is contextually identical
to other VDEV mirrors: you use detach to remove a mirror member.
Interesting, I hadn't read that page. It appears to contradict advice
from other sources (including their own system builds!), maybe they are
just being ultra cautious?
There was a similar discussion on the ZFS list last year which nudged me
in the direction of stripes rather than mirrors:
https://illumos.topicbox.com/groups/zfs/T105516a27e578e84
For my systems (all of which use a stripe of four mirrors), the KVM
performance gain was significant. A single device/mirror could not keep
up with the throughput of the pool.
Cheers,
Ian.
-------------------------------------------
smartos-discuss
Archives: https://www.listbox.com/member/archive/184463/=now
RSS Feed: https://www.listbox.com/member/archive/rss/184463/25769125-55cfbc00
Modify Your Subscription:
https://www.listbox.com/member/?member_id=25769125&id_secret=25769125-7688e9fb
Powered by Listbox: http://www.listbox.com
