NVMe ZNS Makes It Into Linux 5.9 Along With MD RAID Fixes

[WorBlux]

There are pretty serious proposals out there by major manufacturers about hybrid devices, and furthermore, there's no real marginal cost in shipping all CMR drives with this feature. If the OS does nothing with the zones it's not only addressed like CMR, it writes non-overlaping tracks. Random writes, and read-modify operations are just fine w/full performance. If you more space on the drive, you can set it up with SMR zones, you can make zones as SMR, and the drive will shingle the part of the disk and enforce the write only at zone write-pointer restriction. The LBA then grows to reflect the extra space now available.

And I've watch quite a few seminars and talks on the subject and the engineering oriented presenters universally put down host-aware drive. If the host messes up there's a large performance penalty without warnings or errors from the drive.

The bigger problem for hard drives is that it's too little too late, the performance delta vs SSD is so huge. The WD reds you mention are squarely DM-SMR. The didn't report any zbc capability, and WD wouldn't even say they were SMR are first. The only HA-SMR drive I can find is a 2016 8TB Seagate Archive drive (ST8000AS002).

The horrible reality is only the OS can properly guess if the last write to X block was a short term write or a long term one. This is one of the problem with device managed FTL or STL it does not have access to the information it need to make valid action choices. Incorrect choices by FTL/STL lead to write amplification and stalls.
.

On this topic, saw a proposal to expose a tape archive as a glacial storage zoned device. And while the OS can use reasonable heuristics, they aren't always true and some GC and cleanup after the fact is going to be use to correct mistakes.

[oiaohm]

No you won't see a host aware ZNS drive. The zone append command took care of the most tedious part on the host. And the advantage is you can get a fast drive with 1/10th of the onboard DRAM. Host-aware would absolutely kill performance.

The reality is device managed and host aware to attempt to make SMR work with solutions that

Rather you would see and FTL on the host like dm-zoned that will deal with mapping from Logical Block addresses to the actual data location in drive zones.

I would very much like to see it this way from performance.

Plus nobody is actually making zone-aware drives decause they were a bad approach all around, Hybrid-SMR is the better approach (all zones start as conventional addresses and CMR recording patterns, but can be explicitly transitioned to SMR to gain drive space), but still not terribly likely to see wide adoption.

Except there is no such thing as a Hybrid-SMR.


Fun point SMR device managed and Host Aware SMR have something like FTL from SSD its called STL as in a "Shingle Translation Layer".

SMR only come in 3 forms.
1) Device Managed SMR that is where SMR drive pretends to be old school drive with conventional cannot use ZAC/ZBC at all.

https://blocksandfiles.com/2020/04/15/seagate-2-4-and-8tb-barracuda-and-desktop-hdd-smr/

These are in production by samsung.

2)Host Aware SMR.

https://nascompares.com/2020/04/16/your-wd-red-nas-hard-drives-might-be-using-smr-what-you-need-to-know/

These are also in production WD and they are called Host Aware SMR not Hybrid SMR by all harddrive makers. These are the drives that start up they can can take conventiobnal addresses and CMR recording patterns and be by commands transitioned to SMR or have SMR commands use on CMR running sections to trigger storage reorganisations at times its not going to disrupt things.

Yes these are some of the WD red series that gave nice performance problems.

3)Host Managed SMR

https://www.westerndigital.com/products/data-center-drives/ultrastar-dc-hc600-series-hdd

Also in production.

Current day SSD drives are basically flash equal to device managed SMR. Intel is talking about making Host Managed equal in flash. No one I can see is talking about making host aware/hybrid SSD class. But I would not be surprised if we don't see some current day flash drives get firmware updates to come Host Aware class as there re some advantage being Host Aware over being just Device Managed.

The horrible reality is only the OS can properly guess if the last write to X block was a short term write or a long term one. This is one of the problem with device managed FTL or STL it does not have access to the information it need to make valid action choices. Incorrect choices by FTL/STL lead to write amplification and stalls.

Yes host aware SMR are in production and do generate a particular set of problems where you really need to implement all the stuff you need to run the drive in host managed SMR to get ideal performance. Its all to to-do with the lack of information CMR recording patterns provide on the writes being performed. OS has more information at dm-zoned level in the block layer. But at the file system it self at the file system layer has more information again to allow even better informed choices. Yes Windows being dead on this front is why I suspect we will see Host Aware class of SSD its not a good choice but it a better choice for a OS that will up update either its block layer or it file system layer to support zoned storage properly.

[WorBlux]

I would not be too sure of that we have to take ZBC harddrives have host managed and host aware versions. ZNS SSD could have two versions as well in the controllers. Host aware versions would mean non modified OS could use the driver just like they used to. ZNS commands could be used like ZBC on host aware drives for extracting extra information on the drive usage states including possible allowing flash drives to have a user trigger-able block de-fragmentation. So a main stream SSD with ZBC in host aware could claim higher resistance to random slow downs than one without even for a non modified OS like Windows just having to run a manual de-fragmentation tool. SSD vendors are always after to sell higher and higher performance.

The question I really have is are we going to get host aware SSD with ZNS because this will alter how fast we see ZNS SSD in mainstream devices due to the advantages that ZNS usage could give without needing to update OS that much.

F2FS already works with ZNS SSD.

No you won't see a host aware ZNS drive. The zone append command took care of the most tedious part on the host. And the advantage is you can get a fast drive with 1/10th of the onboard DRAM. Host-aware would absolutely kill performance. Rather you would see and FTL on the host like dm-zoned that will deal with mapping from Logical Block addresses to the actual data location in drive zones.

Plus nobody is actually making zone-aware drives decause they were a bad approach all around, Hybrid-SMR is the better approach (all zones start as conventional addresses and CMR recording patterns, but can be explicitly transitioned to SMR to gain drive space), but still not terribly likely to see wide adoption.

But the sad thing is that Windows seems entirely dead on this front anyways. Nobody has done serious file storage on windows for quite a while now, and the filesytems and disk drivers aren't getting any sort of attention beyond the bare minimum anyways.

[oiaohm]

Unless the next version of NVMe specifications make this mandatory it's not going to be included in mainstream devices. I do see the benefits, but the problem is availability in hardware.

I would not be too sure of that we have to take ZBC harddrives have host managed and host aware versions. ZNS SSD could have two versions as well in the controllers. Host aware versions would mean non modified OS could use the driver just like they used to. ZNS commands could be used like ZBC on host aware drives for extracting extra information on the drive usage states including possible allowing flash drives to have a user trigger-able block de-fragmentation. So a main stream SSD with ZBC in host aware could claim higher resistance to random slow downs than one without even for a non modified OS like Windows just having to run a manual de-fragmentation tool. SSD vendors are always after to sell higher and higher performance.

The question I really have is are we going to get host aware SSD with ZNS because this will alter how fast we see ZNS SSD in mainstream devices due to the advantages that ZNS usage could give without needing to update OS that much.

F2FS already works with ZNS SSD.

[starshipeleven]

Thanks for the link. I see more clearly what it is now, but my opinion still stands - it's targeted towards certain use cases in enterprise.

will be useful in a few years when the old enterprise hardware is sold as used.

I got some 20x 600GB SAS drives for like 100 euro a few years ago. They were manufactured in 2013

[starshipeleven]

Does this mean we will be able to use F2FS or UBIFS directly on our SSDs with ZNS?

F2FS is a block device filesystem (as ext4/XFS/Btrfs/whatever), UBI/UBIFS is a raw flash filesystem.

You can already format a SSD with F2FS

[numacross]

Not necessarily, for virtualization is huge as well from simple VMs on your PC to data centers of course, i also expect this to made it into ZFS/BTRFS and helps lots with NVME performance since their allocation schemes are quite complex

Unless the next version of NVMe specifications make this mandatory it's not going to be included in mainstream devices. I do see the benefits, but the problem is availability in hardware.

[jrch2k8]

Thanks for the link. I see more clearly what it is now, but my opinion still stands - it's targeted towards certain use cases in enterprise.

Not necessarily, for virtualization is huge as well from simple VMs on your PC to data centers of course, i also expect this to made it into ZFS/BTRFS and helps lots with NVME performance since their allocation schemes are quite complex

[numacross]

Well, for Enterprise, Cloud and Big Storage make sense since you could have fined grained control over allocations because the controller in most SSD are decent or functional but not even close to good and never will be, because controllers are generic.

Also this is not about how the hardware handle every bit on the flash at hardware level, that is still handled by the hardware but more about allocations, read more here https://zonedstorage.io/introduction/zns/

Thanks for the link. I see more clearly what it is now, but my opinion still stands - it's targeted towards certain use cases in enterprise.

[jrch2k8]

But why? The controller firmware is far more capable and knowledgeable about the exact properties of the underlying flash modules. Apart from very specific use cases I don't think it's going to be beneficial for home or even most enterprise users.

Well, for Enterprise, Cloud and Big Storage make sense since you could have fined grained control over allocations because the controller in most SSD are decent or functional but not even close to good and never will be, because controllers are generic.

Also this is not about how the hardware handle every bit on the flash at hardware level, that is still handled by the hardware but more about allocations, read more here https://zonedstorage.io/introduction/zns/