[zfs-discuss] Optane devices, new Metadata Allocation Class feature

ashford at accs.com ashford at accs.com
Sun Feb 4 11:42:56 EST 2018


On 2018-02-04 02:00, Durval Menezes via zfs-discuss wrote:
> Hello Andreas, Hetz,
> 
> El 04-02-2018 03:57, "Andreas Dilger via zfs-discuss"
> <zfs-discuss at list.zfsonlinux.org> escribió:
> 
>> These would be great ZIL devices, since they are the new 3DXPoint
>> NVRAM (not 3DNAND) so have much higher write endurance and
>> better/consistent IOPS because they don't need to be erased before
>> overwrite like NAND flash.
> 
> Agreed, Optane would be perfect for dedicated ZIL.

I don't believe that any system has two Optane slots, so this ZIL 
probably can't be mirrored for improved reliability.

>> The second thing to consider is that the Metadata Allocation Class
>> feature will be in the 0.8 release, so you could start to allocate
>> all of your metadata on flash to improve performance for scrub, and
>> many workloads. That means you could use excess capacity of a larger
>> flash device for more than just ZIL/L2ARC.  In the MAC issue on
>> GitHub there was some estimation that about 3-5% of the total pool
>> size would be good for an SSD special VDEV to hold all the metadata.
> 
> Another thing I wasn't aware of. Some googling brought up your
> presentation here:
> https://www.eofs.eu/_media/events/lad17/14_andreas_dilger_lad2017-zfs_improvements.pdf
> (slides #13 and #14), which points to
> https://github.com/zfsonlinux/zfs/issues/3779

I hadn't seen this before.  Nice presentation.

> Any info on how this is going (the GitHub page above looks a little
> dead) and whether this is going to be picked up by other ZFS
> implementations, specially FreeBSD?

The work was transferred to a new issue:
https://github.com/zfsonlinux/zfs/pull/5182
It seems that Github doesn't like long issue threads - they sometimes 
stop loading into the browser.

There have been some fundamental changes made to this task.  Originally, 
there was the ability to dedicate VDEVs to Metadata, DDT and small files 
independently.  Now, they're all clumped together.

Peter Ashford

> 
> (PS: changed the 'subject' in order to better reflect these points)
> 
> Cheers,
> --
>    Durval.


More information about the zfs-discuss mailing list