Hi Rao,
On Wed, 28 Mar 2018, Rao Shoaib wrote:
On 03/28/2018 02:39 PM, Mat Martineau wrote:
>
> Hi Rao -
>
> On Wed, 28 Mar 2018, Rao Shoaib wrote:
>
>> I have recently noticed that the group is working on socket level changes
>> to MPTCP.
>>
>> Can a summary of the issues and how the changes will solve them be posted
>> to the list? That would help any new participants as well. If they have
>> already been discussed please point me to the thread.
>
> I proposed moving away from the metasocket architecture last summer:
>
https://lists.01.org/pipermail/mptcp/2017-July/000064.html
>
> When I reviewed the
multipath-tcp.org implementation of MPTCP, my
> assessment was (and is) that using metasockets resulted in more intrusive
> code changes for TCP. The more partitioned approach of KCM and RDS (using
> internal kernel sockets) shows an example of in-kernel features that the
> maintainers have merged. The approach seemed feasible and I proposed it as
> a direction to go.
Thanks for pointing the thread out. That was a proposal we never agreed on
it ?
Yes, I think it's accurate to call it a proposal. The discussion trailed
off without any specific objection to or endorsement of the metasocket
point.
Here is my response to it.
"""""
I agree with the above approach but want to expand on the initial goal.
Our initial goal must be to put a minimal (bare bones) MPTCP implementation
in main stream Linux. That could mean no fancy scheduling schemes, just
simple round robin or just active/standby. Implementing minimal MPTCP
functionality will pretty much expose how main TCP code will be impacted. Any
future work to add features will be confined to changes within MPTCP and
should not be a concern right now. Such an implementation will also be fully
RFC compliant.
I agree with you that upstream folks would want an opt-in option. I am in
favor of a completely different socket family as that would leave tcp socket
code untouched. However we should talk to upstream folks once again.
I should have been more specific, I was referring to the socket layer
changes. Now that I have implemented MPTCP on Linux and understand the design
better I don't think using a separate
I would like us to use as much code as possible from the current
implementation. In fact if for some reason (I don't see one) we can not ship
a minimal implementation than I prefer that we just port the current
implementation and worry about architectural issues later.
In short, I would like our focus to be on putting a minimal MPTCP
implementation in Linux so that we have a stake in the ground. Performance
and Features can come later. That does not mean that the quality of our
implementation is so bad that it is unusable.
Shoaib
""""""""""""""
I was OK using KCM or a different socket family, but nothing else, as I say
using a socket family would only save changes to the socket code.
Ok - appreciate the clarification.
Since our initial goal was not adhered to and I have had more experience,
I would like a technical analysis or better yet an implementation that supports
the said claims. Any mention of upstream should have pointers to the email
threads or should not be used.
I think what people seem to agree on is that an initial upstreamable
implementation will function with a minimal feature set. My understanding
is that we have yet to settle on what the steps are to get to that goal,
and we're working to agree on the steps.
KCM and RDS are completely different, they do
not add options to TCP header or deal with TCP state machine.
I completely agree that they aren't just like MPTCP! And that MPTCP is
harder to integrate because of exactly the reasons you mention. My
intent was to point to them as kernel-internal users of TCP with some
limited design parallels that we can learn from.
I have submitted a patch that provides a complete working implementation
based on the current design. It can be used as a reference to point out how
the proposed solution would be better.
>
>> It will also help if it can be made clear how these changes will
>> expedite implementation of a basic MPTCP implementation. I believe
>> that is the current goal, but my understanding may be incorrect.
>
> I think it will expedite upstreaming in the sense that metasockets would
> never get merged upstream, and that a more partitioned approach with
> dedicated subflow sockets has a reasonable shot. That's my opinion based on
> what I've seen from the maintainers (in terms of mailing list discussions,
> conversations, and what they have merged in the past).
>
> Now, Christoph has raised the topic for more detailed discussion with
> regard to the specifics of the
multipath-tcp.org implementation. I don't
> think anything is a foregone conclusion in terms of this aspect of the
> upstreaming project, you can propose a different approach to discuss.
I want to point out that folks propose a lot of enhancements on Netdev and
IETF with valid uses cases. 99.99% get rejected. There has to be very valid
reasons -- We are not going to build a kitchen sink of options that will make
the design complex and unmaintainable. If a vendor has certain requirements
they need to provide us a list that we evaluate and find out the best way to
implement it if we choose to. We are not there yet and talking about any kind
of options is premature.
To me, the socket-layer changes brought up for discussion are first and
foremost about modifying the design and architecture of MPTCP to reduce
the impact on the TCP code. I haven't seen that anyone is putting current
development effort toward adding bells & whistles - some topics came up
about possible future APIs that might fit well (like exposing additional
fds), but I don't see those as in-scope for initial upstreaming.
Part of what we discussed at the last meeting was using a wiki to better
document what's going on in our MPTCP upstreaming community, I think that
will do a lot to communicate where contributors are putting their current
efforts and also help keep track of ideas for the future.
Regards,
--
Mat Martineau
Intel OTC