LWN: Comments on "Zoned storage and filesystems"

Zone size = hardware erase block size?

DemiMarie — Mon, 19 Jun 2023 14:10:17 +0000

From my perspective, ZNS essentially moves the flash translation layer from the device to the host. Linux filesystems (other than zonefs) running on zoned storage expose a full POSIX API, and that includes random write support. The question is what zone size will yield optimal performance for workloads that are random-write from a userspace perspective.

Zoned storage (SMR) drive end-user opacity

faramir — Mon, 05 Jun 2023 05:34:08 +0000

I tried to figure this out at one point and it seems the issue is:

Drive Managed SMR (consumer drives) vs. Host Managed SMR (enterprise drives).

Consumer SMR (DM) drives don't implement the protocol and who knows what their
firmware is doing underneath. Some people say that they do something like modern SSD
drives (TLC or QLC) which use part of the flash as 'pseudo-SLC' caches in order to provide
a fast write cache which might eventually be moved to TLC/QLC when the SSD gets less busy.
So part of a DM SMR drive might actually not be SMR. Also, the algorithms involved are almost
certainly considered vendor proprietary. It's not even clear it would be reasonable for a DM SMR
drive to attempt to say anything to the Host as it might change it's layout on the fly.

Or at least the above was my take away before I gave up on the subject.

Zone size = hardware erase block size?

adobriyan — Fri, 02 Jun 2023 11:25:09 +0000

> Or are zoned storage devices primarily intended for workloads that bypass the filesystem and handle everything in userspace?

Any application which expects to overwrite data in-place or write data randomly in LBA space is immediately disqualified from using zoned devices and must become small database-like engine (hello, O_DIRECT!) which writes is very precise order (less headache with Zone Append).

So, all classic block/extent filesystems are out. In theory, journalling filesystem may accept zoned device for external journal.

Zone size = hardware erase block size?

DemiMarie — Thu, 01 Jun 2023 22:38:52 +0000

How much does this improve storage performance compared to writing to multiple zones at the same time? Is it worth the decrease in filesystem performance? Or are zoned storage devices primarily intended for workloads that bypass the filesystem and handle everything in userspace?

Zone size = hardware erase block size?

adobriyan — Tue, 30 May 2023 11:53:34 +0000

Internally zones are spread over multiple erase blocks to get better performance on sequential workloads (which constitute 100% of write workloads).
You can guess mapping topology by dividing ZSZE/ZCAP by erase block size.

Zone size = hardware erase block size?

DemiMarie — Mon, 29 May 2023 23:39:39 +0000

Why not have the zone size be equal to the hardware erase block size? That is the smallest value that makes sense, and anything bigger is going to make the lives of filesystem writers unnecessarily harder.

Zoned storage (SMR) drive end-user opacity

Hobart — Fri, 26 May 2023 15:13:41 +0000

Has anyone found a consumer SMR drive that even tells the device owner what the status of its SMR behavior is?

There's been protocol support for years - IIRC Ted T'so had a patch set for extfs to interact with them in a more optimal way - but IDK if any HD manufacturer has actually made that available to the end user.