[LSF/MM TOPIC] Compounding support in CIFS.KO

[LSF/MM TOPIC] Compounding support in CIFS.KO

[ronnie sahlberg]

(Re-sending since typing the right email address is hard)


Proposal:


Compounding support for cifs.ko

In CIFS.KO we are currently adding compounding support for SMB2/3.
As SMB* is a very chatty protocol which likes to do open/query/close a LOT
compounding support to wrap all three operations inside a single round trip
can have a significant performance boost to meta data intensive workloads.

In this talk I will talk about the challenges I faced and what previous design
decisions in cifs.ko that made this very hard and how I refactored the design.

I will also talk about the current status of these patches as well as give a
live demonstration of compounding in action.

regards
Ronnie Sahlberg, RedHat

Re: [LSF/MM TOPIC] Compounding support in CIFS.KO

[Matthew Wilcox]

On Thu, Feb 01, 2018 at 06:23:03AM +1000, ronnie sahlberg wrote:
> Compounding support for cifs.ko
> 
> In CIFS.KO we are currently adding compounding support for SMB2/3.
> As SMB* is a very chatty protocol which likes to do open/query/close a LOT
> compounding support to wrap all three operations inside a single round trip
> can have a significant performance boost to meta data intensive workloads.
> 
> In this talk I will talk about the challenges I faced and what previous design
> decisions in cifs.ko that made this very hard and how I refactored the design.
> 
> I will also talk about the current status of these patches as well as give a
> live demonstration of compounding in action.

I think all of these things are worthwhile and stand to improve
the implementation a lot.  Are they of interest to other filesystem
developers?  Are there VFS changes which would help make compounding
easier?  Or is this only of interest to other CIFS people, in which case
this might not be the best use of a slot at LSFMM ...

Re: [LSF/MM TOPIC] Compounding support in CIFS.KO

[Chuck Lever]

> On Jan 31, 2018, at 5:03 PM, Matthew Wilcox <willy@infradead.org> wrote:
> 
> On Thu, Feb 01, 2018 at 06:23:03AM +1000, ronnie sahlberg wrote:
>> Compounding support for cifs.ko
>> 
>> In CIFS.KO we are currently adding compounding support for SMB2/3.
>> As SMB* is a very chatty protocol which likes to do open/query/close a LOT
>> compounding support to wrap all three operations inside a single round trip
>> can have a significant performance boost to meta data intensive workloads.
>> 
>> In this talk I will talk about the challenges I faced and what previous design
>> decisions in cifs.ko that made this very hard and how I refactored the design.
>> 
>> I will also talk about the current status of these patches as well as give a
>> live demonstration of compounding in action.
> 
> I think all of these things are worthwhile and stand to improve
> the implementation a lot.  Are they of interest to other filesystem
> developers?  Are there VFS changes which would help make compounding
> easier?  Or is this only of interest to other CIFS people, in which case
> this might not be the best use of a slot at LSFMM ...

This is in my area of interest also.

However, my understanding of LSF/MM is that this is not a "stand up
and give a status or implementation experience report" kind of event.
It's more of a "I have this specific itch that I need to get some
help with." None of the above proposal strikes me as very specific,
and a CIFS-centric talk would likely be unappealing to even most of
the attendees of the Filesystems track.

Can the proposed topic be rephrased as a focused discussion of what
work needs to be moved forward, what decisions need to be made, and
so on? Otherwise, we should move this topic to a network filesystems
BoF or to the corridor track, IMHO.


--
Chuck Lever

Re: [LSF/MM TOPIC] Compounding support in CIFS.KO

[ronnie sahlberg]

On Thu, Feb 1, 2018 at 8:13 AM, Chuck Lever <chuck.lever@oracle.com> wrote:
>
>
>> On Jan 31, 2018, at 5:03 PM, Matthew Wilcox <willy@infradead.org> wrote:
>>
>> On Thu, Feb 01, 2018 at 06:23:03AM +1000, ronnie sahlberg wrote:
>>> Compounding support for cifs.ko
>>>
>>> In CIFS.KO we are currently adding compounding support for SMB2/3.
>>> As SMB* is a very chatty protocol which likes to do open/query/close a LOT
>>> compounding support to wrap all three operations inside a single round trip
>>> can have a significant performance boost to meta data intensive workloads.
>>>
>>> In this talk I will talk about the challenges I faced and what previous design
>>> decisions in cifs.ko that made this very hard and how I refactored the design.
>>>
>>> I will also talk about the current status of these patches as well as give a
>>> live demonstration of compounding in action.
>>
>> I think all of these things are worthwhile and stand to improve
>> the implementation a lot.  Are they of interest to other filesystem
>> developers?  Are there VFS changes which would help make compounding
>> easier?  Or is this only of interest to other CIFS people, in which case
>> this might not be the best use of a slot at LSFMM ...
>
> This is in my area of interest also.
>
> However, my understanding of LSF/MM is that this is not a "stand up
> and give a status or implementation experience report" kind of event.
> It's more of a "I have this specific itch that I need to get some
> help with." None of the above proposal strikes me as very specific,
> and a CIFS-centric talk would likely be unappealing to even most of
> the attendees of the Filesystems track.
>
> Can the proposed topic be rephrased as a focused discussion of what
> work needs to be moved forward, what decisions need to be made, and
> so on? Otherwise, we should move this topic to a network filesystems
> BoF or to the corridor track, IMHO.
>

I think you are right.
The topic of cifs compounding is only really interesting to cifs folks and
not many others.
While other filesystems such as NFSv4 has compounding too, the way compounding
works in cifs is different enough that there is little meaningful
overlap between the two.

It is therefore probably not suitable for lsf/mm.


Please disregard this proposal.
It might be more suitable to just have ad-hoc informal talks between
the cifs folks at the event
than a scheduled topic.


regards
Ronnie Sahlberg

Re: [LSF/MM TOPIC] Compounding support in CIFS.KO

[J. Bruce Fields]

On Thu, Feb 01, 2018 at 08:22:03AM +1000, ronnie sahlberg wrote:
> I think you are right.  The topic of cifs compounding is only really
> interesting to cifs folks and not many others.  While other
> filesystems such as NFSv4 has compounding too, the way compounding
> works in cifs is different enough that there is little meaningful
> overlap between the two.

I'm trying to remember what people have found to be the obstacles to
more aggressive use of compounds in NFS.  (Things like opening a file
and doing IO to it in one compound.)  Might it be more practical with
some VFS changes?  Are there some shared CIFS/NFS issues there?

Anyway, the hallway track is fine....

--b.

Re: [LSF/MM TOPIC] Compounding support in CIFS.KO

[Steve French]

On Wed, Jan 31, 2018 at 8:40 PM, J. Bruce Fields <bfields@fieldses.org> wrote:
> On Thu, Feb 01, 2018 at 08:22:03AM +1000, ronnie sahlberg wrote:
>> I think you are right.  The topic of cifs compounding is only really
>> interesting to cifs folks and not many others.  While other
>> filesystems such as NFSv4 has compounding too, the way compounding
>> works in cifs is different enough that there is little meaningful
>> overlap between the two.
>
> I'm trying to remember what people have found to be the obstacles to
> more aggressive use of compounds in NFS.  (Things like opening a file
> and doing IO to it in one compound.)  Might it be more practical with
> some VFS changes?  Are there some shared CIFS/NFS issues there?

Very likely - things that results in the equivalent of getattr/open/getattr
for example.  Should be possible to find examples where multiple
VFS calls could be optionally compressed into one (for network file systems
where the individual calls are expensive to send independently).

> Anyway, the hallway track is fine....



-- 
Thanks,

Steve