To my understanding, a resilver follows the logical order of the blocks in
the ZIO? That means that it will follow (as you suggest) the fragmentation
of the pool. Maybe during a resilver we can read everything 'fragmented',
then write the resilver actions to the ZIL/SLOG, and in a separate action
(thread?) write to the resilvering disks sequentially (if there is enough
space).
This will be greatly beneficial for SMR disks because they can do a fair
amount of random reads, but they lag behind tremendously with random writes
(order of magnitude slower than reads).

I think it may be to high of a risk to write to the SLOG and then
re-reading it for writing to the resilvering disks.
But the benefits for the resilver process of SMR disks might justify the
risk.

- Thijs

2016-02-08 22:07 GMT+01:00 Matthew Ahrens <[email protected]>:

>
>
> On Sun, Nov 1, 2015 at 11:44 PM, Thijs Cramer <[email protected]>
> wrote:
>
>> Hi guys,
>>
>> Recently in a discussion online, i discussed the new Seagate 8TB SMR
>> disks in combination with ZFS.
>> A big point that was undecided was: Can a SLOG help during rebuild of a
>> bit SMR disk based array?
>>
>> I know it's more risky to use the SLOG for resilver activities, but
>> rebuilding sequentially could benefit these disks insanely. The problem
>> with these disks is that under heavy load, the disk because very slow on
>> random writes because of the SMR rewrites the disk has to do.
>>
>> Can anyone answer whether it's currently beneficial to have a SLOG on an
>> SMR disk array, or answer whether it's easy to adapt the code a bit to make
>> this happen?
>>
>>
> Currently, the ZIL (and thus log devices / SLOG) is not involved in
> resilvering.  Therefore adding a SLOG device will have no impact on
> resilver speed.  (Though I guess there could be secondary effects if adding
> the SLOG reduces fragmentation in the main devices.)
>
> If you have a specific proposal for how the ZIL / SLOG could be used to
> accelerate resilvering, I can give you feedback on its feasibility.
>
> --matt
>
>
_______________________________________________
developer mailing list
[email protected]
http://lists.open-zfs.org/mailman/listinfo/developer

Reply via email to