[LSF/MM/BPF BoF] Session for Zoned Storage 2023

[LSF/MM/BPF BoF] Session for Zoned Storage 2023

[Viacheslav Dubeyko]

Hello,

I would like to suggest to have ZNS related session again. I think we have a lot of to discuss.
As far as I can see, I have two topics for discussion. And I believe there are multiple other
topics too.

How everybody feels to have a room for ZNS related discussion?

Thanks,
Slava.

Re: [LSF/MM/BPF BoF] Session for Zoned Storage 2023

[Luis Chamberlain]

On Fri, Jan 06, 2023 at 11:17:19AM -0800, Viacheslav Dubeyko wrote:
> Hello,
> 
> As far as I can see, I have two topics for discussion.

What's that?

  Luis

Re: [LSF/MM/BPF BoF] Session for Zoned Storage 2023

[Viacheslav Dubeyko]

> On Jan 6, 2023, at 11:18 AM, Luis Chamberlain <mcgrof@kernel.org> wrote:
> 
> On Fri, Jan 06, 2023 at 11:17:19AM -0800, Viacheslav Dubeyko wrote:
>> Hello,
>> 
>> As far as I can see, I have two topics for discussion.
> 
> What's that?

I am going to share these topics in separate emails. :)

(1) I am going to share SSDFS patchset soon. And topic is:
SSDFS + ZNS SSD: deterministic architecture decreasing TCO cost of data infrastructure.

(2) Second topic is:
How to achieve better lifetime and performance of caching layer with ZNS SSD?

Thanks,
Slava.

Re: [External] [LSF/MM/BPF BoF] Session for Zoned Storage 2023

[Viacheslav A.Dubeyko]

CC: LSF/MM/BPF mailing list. Sorry, missed the list.

> On Jan 6, 2023, at 11:17 AM, Viacheslav Dubeyko <slava@dubeyko.com> wrote:
> 
> Hello,
> 
> I would like to suggest to have ZNS related session again. I think we have a lot of to discuss.
> As far as I can see, I have two topics for discussion. And I believe there are multiple other
> topics too.
> 
> How everybody feels to have a room for ZNS related discussion?
> 
> Thanks,
> Slava.

Re: [External] [LSF/MM/BPF BoF] Session for Zoned Storage 2023

[Viacheslav A.Dubeyko]

> On Jan 6, 2023, at 11:30 AM, Viacheslav Dubeyko <slava@dubeyko.com> wrote:
> 
> 
> 
>> On Jan 6, 2023, at 11:18 AM, Luis Chamberlain <mcgrof@kernel.org> wrote:
>> 
>> On Fri, Jan 06, 2023 at 11:17:19AM -0800, Viacheslav Dubeyko wrote:
>>> Hello,
>>> 
>>> As far as I can see, I have two topics for discussion.
>> 
>> What's that?
> 
> I am going to share these topics in separate emails. :)
> 
> (1) I am going to share SSDFS patchset soon. And topic is:
> SSDFS + ZNS SSD: deterministic architecture decreasing TCO cost of data infrastructure.
> 
> (2) Second topic is:
> How to achieve better lifetime and performance of caching layer with ZNS SSD?
> 

I think we can consider such discussions:
(1) I assume that we still need to discuss PO2 zone sizes?
(2) Status of ZNS SSD support in F2FS, btrfs (maybe, bcachefs and other file systems)
(3) Any news from ZoneFS (+ ZenFS maybe)?
(4) New ZNS standard features that we need to support on block layer + FS levels?
(5) ZNS drive emulation + additional testing features?
(6) ZNS + computational drive? What new features would we like to see from ZNS SSD?
(7) ZNS + CXL: does it make sense?

Thanks,
Slava

Re: [External] [LSF/MM/BPF BoF] Session for Zoned Storage 2023

[Bart Van Assche]

On 1/9/23 07:33, Javier González wrote:
> On 06.01.2023 17:56, Viacheslav A.Dubeyko wrote:
[ ... ]
> As a general comment, do we want to talk about ZNS alone, or zoned
> storage in general? We have of course SMR, but also zoned UFS with
> actual use-cases.

I propose to broaden the conversation from ZNS to ZNS (NVMe) + ZBC (SCSI).

>> I think we can consider such discussions:
>> (1) I assume that we still need to discuss PO2 zone sizes?
> 
> For this discussion to move forward, we need users rather than vendors
> talking about the need. If someone is willing to drive this discussion,
> then it makes sense. I do not believe we will make progress otherwise.

In JEDEC meetings I hear that UFS vendors strongly request support for 
zone sizes that are not a power of two. The JEDEC UFS committee is 
currently busy with requesting an MoU from T10 for permission to base 
JEDEC standards on ZBC. We plan to finalize the ZUFS (zoned UFS) 
specification once that MoU has been established (probably later this 
spring).

An additional topic I want to talk about is support for queue depths > 1 
for sequential write required zone type. I plan to post patches soon 
(later this week).

Thanks,

Bart.

Re: [External] [LSF/MM/BPF BoF] Session for Zoned Storage 2023

[Viacheslav A.Dubeyko]

> On Jan 9, 2023, at 7:33 AM, Javier González <javier.gonz@samsung.com> wrote:
> 

<skipped>

>>> 
>>> (1) I am going to share SSDFS patchset soon. And topic is:
>>> SSDFS + ZNS SSD: deterministic architecture decreasing TCO cost of data infrastructure.
> 
> 
> Would be good to see the patches before LSF/MM/BPF.
> 

I am making code cleanup now. I am expecting to share patches in two weeks.

> I saw your talk at Plumbers. Do you think you have more data to share
> too? Maybe even a comparisson with btrfs in terms of WAF and Space Amp?
> 

I am working to share more data. So, I should have more details.
I have data for btrfs already. Do you mean that you would like to see comparison
btrfs + compression vs. ssdfs? By the way, I am using my own methodology
to estimate WAF and space amplification. What methodology do you have in mind?
Maybe, I could improve mine. :)

<skipped>

>>> 
>> 
>> I think we can consider such discussions:
>> (1) I assume that we still need to discuss PO2 zone sizes?
> 
> For this discussion to move forward, we need users rather than vendors
> talking about the need. If someone is willing to drive this discussion,
> then it makes sense. I do not believe we will make progress otherwise.
> 

As part of ByteDance, I am on user side now. :) So, let me have some internal
discussion and to summarize vision(s) on our side. I believe that, maybe, it makes
sense to summarize a list of pros and cons and to have something like analysis or
brainstorming here.

<skipped>

> 
>> (4) New ZNS standard features that we need to support on block layer + FS levels?
> 
> Do you have any concrete examples in mind?
> 

My point here that we could summarize:
(1) what features already implemented and supported,
(2) what features are under implementation and what is progress,
(3) what features need to be implemented yet.

Have we implemented everything already? :)

>> (5) ZNS drive emulation + additional testing features?
> 
> Is this QEMU alone or do you have other ideas in mind?
> 

My point is the same here. Let’s summarize how reasonably good is emulation now.
Do we need to support the emulation of any additional features?
And we can talk not only about QEMU.

Thanks,
Slava.
 

Re: [External] [LSF/MM/BPF BoF] Session for Zoned Storage 2023

[Damien Le Moal]

On 1/10/23 04:11, Viacheslav A.Dubeyko wrote:
> 
>> On Jan 9, 2023, at 7:33 AM, Javier González <javier.gonz@samsung.com> wrote:
>>
> 
> <skipped>
> 
>>>>
>>>> (1) I am going to share SSDFS patchset soon. And topic is:
>>>> SSDFS + ZNS SSD: deterministic architecture decreasing TCO cost of data infrastructure.
>>
>>
>> Would be good to see the patches before LSF/MM/BPF.
>>
> 
> I am making code cleanup now. I am expecting to share patches in two weeks.
> 
>> I saw your talk at Plumbers. Do you think you have more data to share
>> too? Maybe even a comparisson with btrfs in terms of WAF and Space Amp?
>>
> 
> I am working to share more data. So, I should have more details.
> I have data for btrfs already. Do you mean that you would like to see comparison
> btrfs + compression vs. ssdfs? By the way, I am using my own methodology
> to estimate WAF and space amplification. What methodology do you have in mind?
> Maybe, I could improve mine. :)
> 
> <skipped>
> 
>>>>
>>>
>>> I think we can consider such discussions:
>>> (1) I assume that we still need to discuss PO2 zone sizes?
>>
>> For this discussion to move forward, we need users rather than vendors
>> talking about the need. If someone is willing to drive this discussion,
>> then it makes sense. I do not believe we will make progress otherwise.
>>
> 
> As part of ByteDance, I am on user side now. :) So, let me have some internal
> discussion and to summarize vision(s) on our side. I believe that, maybe, it makes
> sense to summarize a list of pros and cons and to have something like analysis or
> brainstorming here.
> 
> <skipped>
> 
>>
>>> (4) New ZNS standard features that we need to support on block layer + FS levels?
>>
>> Do you have any concrete examples in mind?
>>
> 
> My point here that we could summarize:
> (1) what features already implemented and supported,
> (2) what features are under implementation and what is progress,
> (3) what features need to be implemented yet.
> 
> Have we implemented everything already? :)

Standards are full of features that are not useful in a general purpose
system. So we likely never will implement everything. We never did for
SCSI and ATA and never will either.

> 
>>> (5) ZNS drive emulation + additional testing features?
>>
>> Is this QEMU alone or do you have other ideas in mind?
>>
> 
> My point is the same here. Let’s summarize how reasonably good is emulation now.
> Do we need to support the emulation of any additional features?
> And we can talk not only about QEMU.
> 
> Thanks,
> Slava.
>  

-- 
Damien Le Moal
Western Digital Research

Re: [External] [LSF/MM/BPF BoF] Session for Zoned Storage 2023

[Jens Axboe]

>> My point here that we could summarize:
>> (1) what features already implemented and supported,
>> (2) what features are under implementation and what is progress,
>> (3) what features need to be implemented yet.
>>
>> Have we implemented everything already? :)
> 
> Standards are full of features that are not useful in a general purpose
> system. So we likely never will implement everything. We never did for
> SCSI and ATA and never will either.
Indeed, and that's a very important point. Some people read specs and
find things that aren't in the Linux driver (any spec, not a specific
one), and think they need to be added. No. We only add them if they make
sense, both in terms of use cases, but also as long as they can get
implemented cleanly. Parts of basically any spec is garbage and don't
necessarily fit within the given subsystem either.

The above would make me worried about patches coming from anyone with
that mindset.

-- 
Jens Axboe

Re: [External] [LSF/MM/BPF BoF] Session for Zoned Storage 2023

[Viacheslav A.Dubeyko]

> On Jan 9, 2023, at 3:00 PM, Jens Axboe <axboe@kernel.dk> wrote:
> 
>>> My point here that we could summarize:
>>> (1) what features already implemented and supported,
>>> (2) what features are under implementation and what is progress,
>>> (3) what features need to be implemented yet.
>>> 
>>> Have we implemented everything already? :)
>> 
>> Standards are full of features that are not useful in a general purpose
>> system. So we likely never will implement everything. We never did for
>> SCSI and ATA and never will either.
> Indeed, and that's a very important point. Some people read specs and
> find things that aren't in the Linux driver (any spec, not a specific
> one), and think they need to be added. No. We only add them if they make
> sense, both in terms of use cases, but also as long as they can get
> implemented cleanly. Parts of basically any spec is garbage and don't
> necessarily fit within the given subsystem either.
> 
> The above would make me worried about patches coming from anyone with
> that mindset.
> 

OK. We already have discussion about garbage in spec. :)
So, what would we like finally implement and what never makes sense to do?
Should we identify really important stuff for implementation?

Thanks,
Slava.

Re: [External] [LSF/MM/BPF BoF] Session for Zoned Storage 2023

[Damien Le Moal]

On 1/10/23 08:20, Viacheslav A.Dubeyko wrote:
> 
> 
>> On Jan 9, 2023, at 3:00 PM, Jens Axboe <axboe@kernel.dk> wrote:
>>
>>>> My point here that we could summarize:
>>>> (1) what features already implemented and supported,
>>>> (2) what features are under implementation and what is progress,
>>>> (3) what features need to be implemented yet.
>>>>
>>>> Have we implemented everything already? :)
>>>
>>> Standards are full of features that are not useful in a general purpose
>>> system. So we likely never will implement everything. We never did for
>>> SCSI and ATA and never will either.
>> Indeed, and that's a very important point. Some people read specs and
>> find things that aren't in the Linux driver (any spec, not a specific
>> one), and think they need to be added. No. We only add them if they make
>> sense, both in terms of use cases, but also as long as they can get
>> implemented cleanly. Parts of basically any spec is garbage and don't
>> necessarily fit within the given subsystem either.
>>
>> The above would make me worried about patches coming from anyone with
>> that mindset.
>>
> 
> OK. We already have discussion about garbage in spec. :)
> So, what would we like finally implement and what never makes sense to do?
> Should we identify really important stuff for implementation?

If users need and request something that is not currently supported, we
can discuss implementing it. Until such user-driven request is raised, I
personally do not see any point in discussing anything.

> 
> Thanks,
> Slava.
> 
> 
> 

-- 
Damien Le Moal
Western Digital Research

Re: [External] [LSF/MM/BPF BoF] Session for Zoned Storage 2023

[Jens Axboe]

On 1/9/23 4:20?PM, Viacheslav A.Dubeyko wrote:
> 
> 
>> On Jan 9, 2023, at 3:00 PM, Jens Axboe <axboe@kernel.dk> wrote:
>>
>>>> My point here that we could summarize:
>>>> (1) what features already implemented and supported,
>>>> (2) what features are under implementation and what is progress,
>>>> (3) what features need to be implemented yet.
>>>>
>>>> Have we implemented everything already? :)
>>>
>>> Standards are full of features that are not useful in a general purpose
>>> system. So we likely never will implement everything. We never did for
>>> SCSI and ATA and never will either.
>> Indeed, and that's a very important point. Some people read specs and
>> find things that aren't in the Linux driver (any spec, not a specific
>> one), and think they need to be added. No. We only add them if they make
>> sense, both in terms of use cases, but also as long as they can get
>> implemented cleanly. Parts of basically any spec is garbage and don't
>> necessarily fit within the given subsystem either.
>>
>> The above would make me worried about patches coming from anyone with
>> that mindset.
>>
> 
> OK. We already have discussion about garbage in spec. :)
> So, what would we like finally implement and what never makes sense to do?
> Should we identify really important stuff for implementation?

Well if you did have that discussion, then it seemed you got nothing
from it. Because asking that kind of question is EXACTLY what I'm saying
is the opposite of what should be done. If there's a demand for a
feature, then it can be looked at and ultimately implemented if it makes
sense. You're still talking about proactively finding features and
implementing them "just in case they are needed", which is very much the
opposite and wrong approach, and how any kind of software ends up being
bloated, slow, and buggy/useless.

-- 
Jens Axboe

Re: [External] [LSF/MM/BPF BoF] Session for Zoned Storage 2023

[Viacheslav A.Dubeyko]

> On Jan 9, 2023, at 5:09 PM, Jens Axboe <axboe@kernel.dk> wrote:
> 
> On 1/9/23 4:20?PM, Viacheslav A.Dubeyko wrote:
>> 
>> 
>>> On Jan 9, 2023, at 3:00 PM, Jens Axboe <axboe@kernel.dk> wrote:
>>> 
>>>>> My point here that we could summarize:
>>>>> (1) what features already implemented and supported,
>>>>> (2) what features are under implementation and what is progress,
>>>>> (3) what features need to be implemented yet.
>>>>> 
>>>>> Have we implemented everything already? :)
>>>> 
>>>> Standards are full of features that are not useful in a general purpose
>>>> system. So we likely never will implement everything. We never did for
>>>> SCSI and ATA and never will either.
>>> Indeed, and that's a very important point. Some people read specs and
>>> find things that aren't in the Linux driver (any spec, not a specific
>>> one), and think they need to be added. No. We only add them if they make
>>> sense, both in terms of use cases, but also as long as they can get
>>> implemented cleanly. Parts of basically any spec is garbage and don't
>>> necessarily fit within the given subsystem either.
>>> 
>>> The above would make me worried about patches coming from anyone with
>>> that mindset.
>>> 
>> 
>> OK. We already have discussion about garbage in spec. :)
>> So, what would we like finally implement and what never makes sense to do?
>> Should we identify really important stuff for implementation?
> 
> Well if you did have that discussion, then it seemed you got nothing
> from it. Because asking that kind of question is EXACTLY what I'm saying
> is the opposite of what should be done. If there's a demand for a
> feature, then it can be looked at and ultimately implemented if it makes
> sense. You're still talking about proactively finding features and
> implementing them "just in case they are needed", which is very much the
> opposite and wrong approach, and how any kind of software ends up being
> bloated, slow, and buggy/useless.
> 

I simply tried to suggest some space for this discussion and nothing more.
If all important features have been implemented already and nobody would
like to discuss new feature(s), then we can simply exclude this topic from the list.

If you would like to say that I am a reason of slow software, then I take this credit. :)

Thanks,
Slava.

Re: [External] [LSF/MM/BPF BoF] Session for Zoned Storage 2023

[Jens Axboe]

On 1/9/23 6:39?PM, Viacheslav A.Dubeyko wrote:
> 
> 
>> On Jan 9, 2023, at 5:09 PM, Jens Axboe <axboe@kernel.dk> wrote:
>>
>> On 1/9/23 4:20?PM, Viacheslav A.Dubeyko wrote:
>>>
>>>
>>>> On Jan 9, 2023, at 3:00 PM, Jens Axboe <axboe@kernel.dk> wrote:
>>>>
>>>>>> My point here that we could summarize:
>>>>>> (1) what features already implemented and supported,
>>>>>> (2) what features are under implementation and what is progress,
>>>>>> (3) what features need to be implemented yet.
>>>>>>
>>>>>> Have we implemented everything already? :)
>>>>>
>>>>> Standards are full of features that are not useful in a general purpose
>>>>> system. So we likely never will implement everything. We never did for
>>>>> SCSI and ATA and never will either.
>>>> Indeed, and that's a very important point. Some people read specs and
>>>> find things that aren't in the Linux driver (any spec, not a specific
>>>> one), and think they need to be added. No. We only add them if they make
>>>> sense, both in terms of use cases, but also as long as they can get
>>>> implemented cleanly. Parts of basically any spec is garbage and don't
>>>> necessarily fit within the given subsystem either.
>>>>
>>>> The above would make me worried about patches coming from anyone with
>>>> that mindset.
>>>>
>>>
>>> OK. We already have discussion about garbage in spec. :)
>>> So, what would we like finally implement and what never makes sense to do?
>>> Should we identify really important stuff for implementation?
>>
>> Well if you did have that discussion, then it seemed you got nothing
>> from it. Because asking that kind of question is EXACTLY what I'm saying
>> is the opposite of what should be done. If there's a demand for a
>> feature, then it can be looked at and ultimately implemented if it makes
>> sense. You're still talking about proactively finding features and
>> implementing them "just in case they are needed", which is very much the
>> opposite and wrong approach, and how any kind of software ends up being
>> bloated, slow, and buggy/useless.
>>
> 
> I simply tried to suggest some space for this discussion and nothing
> more. If all important features have been implemented already and
> nobody would like to discuss new feature(s), then we can simply
> exclude this topic from the list.

If something is missing and there's a bof/session, then someone will
bring it up. Fishing for things to implement is not a good idea.

-- 
Jens Axboe

Re: [External] [LSF/MM/BPF BoF] Session for Zoned Storage 2023

[Ming Lei]

On Fri, Jan 06, 2023 at 05:56:24PM -0800, Viacheslav A.Dubeyko wrote:
> 
> 
> > On Jan 6, 2023, at 11:30 AM, Viacheslav Dubeyko <slava@dubeyko.com> wrote:
> > 
> > 
> > 
> >> On Jan 6, 2023, at 11:18 AM, Luis Chamberlain <mcgrof@kernel.org> wrote:
> >> 
> >> On Fri, Jan 06, 2023 at 11:17:19AM -0800, Viacheslav Dubeyko wrote:
> >>> Hello,
> >>> 
> >>> As far as I can see, I have two topics for discussion.
> >> 
> >> What's that?
> > 
> > I am going to share these topics in separate emails. :)
> > 
> > (1) I am going to share SSDFS patchset soon. And topic is:
> > SSDFS + ZNS SSD: deterministic architecture decreasing TCO cost of data infrastructure.
> > 
> > (2) Second topic is:
> > How to achieve better lifetime and performance of caching layer with ZNS SSD?
> > 
> 
> I think we can consider such discussions:
> (1) I assume that we still need to discuss PO2 zone sizes?
> (2) Status of ZNS SSD support in F2FS, btrfs (maybe, bcachefs and other file systems)
> (3) Any news from ZoneFS (+ ZenFS maybe)?
> (4) New ZNS standard features that we need to support on block layer + FS levels?
> (5) ZNS drive emulation + additional testing features?

ZNS drive emulation can be done as one ublk target(userspace
implementation with ublk driver extension), and it was discussed before:

https://github.com/ming1/ubdsrv/pull/28


Thanks, 
Ming

Re: [External] [LSF/MM/BPF BoF] Session for Zoned Storage 2023

[Andreas Hindborg]

Hi Ming, Matias,

Matias Bjørling <Matias.Bjorling@wdc.com> writes:

<snip>

>> >
>> > I think we can consider such discussions:
>> > (1) I assume that we still need to discuss PO2 zone sizes?
>> > (2) Status of ZNS SSD support in F2FS, btrfs (maybe, bcachefs and
>> > other file systems)
>> > (3) Any news from ZoneFS (+ ZenFS maybe)?
>> > (4) New ZNS standard features that we need to support on block layer + FS
>> levels?
>> > (5) ZNS drive emulation + additional testing features?
>> 
>> ZNS drive emulation can be done as one ublk target(userspace implementation
>> with ublk driver extension), and it was discussed before:
>> 
>> https://github.com/ming1/ubdsrv/pull/28
>> 
>> 
>> Thanks,
>> Ming
>
> Hi Ming, 
>
> Andreas recently transitioned from WD to another company. I am not sure if he'll continue this specific work, if he isn’t, my team will pick it up. It already looks very promising.
>
> Thanks, Matias

I joined Samsung recently, and I plan to continue the work on adding
zoned storage support to ublk in that context.

I will send the patches to the relevant lists when they get further
along. I hope your team will evaluate and provide feedback Matias :)

Best regards,
Andreas