Re: [multipathtcp] MPTCP implementation feedback for RFC6824bis

[V Anil Kumar <[e-mail]>]

Hi Christoph,

On 12/13/19 11:54 PM, Christoph Paasch  <[e-mail]> wrote: 
> 
> On 13/12/19 - 23:41:25, V Anil Kumar wrote:
> > Hi Christoph,
> > 
> > Thanks again for your reply. My response is given inline.
> > 
> > On 12/13/19 09:59 PM, Christoph Paasch <[e-mail]> wrote: 
> > > 
> > > 
> > > 
> > > 
> > > 
> > > 
> > > Hello,
> > > 
> > > 
> > > 
> > > > On Dec 12, 2019, at 9:53 PM, V Anil Kumar <[e-mail]> wrote:
> > > > 
> > > > 
> > > 
> > > 
> > > > 
> > > > Hi Christoph,
> > > > 
> > > > Thanks for your reply. Please see inline.
> > > > 
> > > > On 12/12/19 12:52 AM, Christoph Paasch <[e-mail]> wrote: 
> > > > > 
> > > > > 
> > > > > 
> > > > > 
> > > > > 
> > > > > Hello,
> > > > > 
> > > > > 
> > > > > 
> > > > > 
> > > > > > 
> > > > > > On Dec 10, 2019, at 12:04 PM, V Anil Kumar
> > > > > > <[e-mail](javascript:main.compose()> wrote:
> > > > > > 
> > > > > > 
> > > > > > Hi Alan, 
> > > > > > 
> > > > > > 
> > > > > > 
> > > > > > Please see inline.
> > > > > > 
> > > > > > On 12/06/19 09:28 PM,Alan
> > > > > > Ford<[e-mail](javascript:main.compose()> wrote: 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > Hi all,
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > Following on from the discussion of implementation feedback with
> > > > > > > Christoph, I propose the following edits to RFC6824bis - which
> > > > > > > is currently in AUTH48 - as clarifications.
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > ADs, please can you confirm you consider these edits
> > > > > > > sufficiently editorial to fit into AUTH48.
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > WG participants, please speak up if you have any concerns.
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > Edit 1, clarifying reliability of MP_CAPABLE
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > Change the sentence reading:
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > The SYN with MP_CAPABLE occupies the first octet of data
> > > > > > > sequence space, although this does not need to be acknowledged
> > > > > > > at the connection level until the first data is sent (see
> > > > > > > Section 3.3).
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > To:
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > The SYN with MP_CAPABLE occupies the first octet of data
> > > > > > > sequence space, and this MUST be acknowledged at the connection
> > > > > > > level at or before the time the first data is sent or received
> > > > > > > (see Section 3.3).
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > Change the sentence reading:
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > If B has data to send first, then the reliable delivery of the
> > > > > > > ACK + MP_CAPABLE can be inferred by the receipt of this data
> > > > > > > with an MPTCP Data Sequence Signal (DSS) option (Section 3.3). 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > To:
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > If B has data to send first, then the reliable delivery of the
> > > > > > > ACK + MP_CAPABLE is ensured by the receipt of this data with an
> > > > > > > MPTCP Data Sequence Signal (DSS) option (Section 3.3)
> > > > > > > containing a DATA_ACK for the MP_CAPABLE (which is the first
> > > > > > > octet of the data sequence space).
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > In my personal opinion either one of these edits would be
> > > > > > > sufficient for making the point, however clearly this has caused
> > > > > > > some confusion amongst the implementor community so making both
> > > > > > > these changes should make it absolutely clear as to the expected
> > > > > > > behaviour here.
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > Edit 2, mapping constraint
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > Change the sentence reading:
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > A Data Sequence Mapping does not need to be included in every
> > > > > > > MPTCP packet, as long as the subflow sequence space in that
> > > > > > > packet is covered by a mapping known at the receiver.
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > To:
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > A Data Sequence Mapping MUST appear on a TCP segment which is
> > > > > > > covered by the mapping. It does not need to be included in
> > > > > > > every MPTCP packet, as long as the subflow sequence space in
> > > > > > > that packet is covered by a mapping known at the receiver.
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > 
> > > > > > As far as I understand, the proposed change introduces a “MUST” to
> > > > > > insist that the map in a segment must cover at least some data in
> > > > > > the segment. But the document does not talk anything about the
> > > > > > rational behind it. I guess it is purely an 
> > > > > > 
> > > > > > ease of implementation?
> > > > > > 
> > > > > > 
> > > > > 
> > > > > 
> > > > > For two reasons:
> > > > > 
> > > > > 1. Ease of implementation 2. If an implementation tries to
> > > > > "remember" early mappings, it is not clear how many of these an
> > > > > implementation can hold. Thus, the sender does not know how many
> > > > > early mappings he can send. So, it is hard for a sender to do the
> > > > > right thing.
> > > > > 
> > > > > 
> > > > > > 
> > > > > > 
> > > > > > 
> > > > > > I think the design/format of the Data Sequence Mapping permits the
> > > > > > map to stand independent of the data being carried in a segment.
> > > > > > So, as long as an implementation is willing to deal with the
> > > > > > complexity of storing and processing late and early mappings (with
> > > > > > respect to the data arrival), it could be permitted provided that
> > > > > > the received map is for an in-window data.
> > > > > > 
> > > > > > 
> > > > > 
> > > > > 
> > > > > What is the concrete use-case for such early mappings? What are the
> > > > > benefits of it? I think that if we want to enable such
> > > > > implementation-complexity, we need a compelling use-case with a big
> > > > > benefit.
> > > > > 
> > > > > 
> > > > > 
> > > > > 
> > > > 
> > > > Consider a case where a MPTCP connection consists of two subflows, and
> > > > the data segments are scheduled for transmission in the order shown
> > > > below below. 
> > > > 
> > > > 
> > > > 
> > > > 
> > > > 
> > > > 
> > > > Subflow-1: segment-1 segment-3 segment-5 segment-6 
> > > > 
> > > > bytes:1-100 bytes:201-300 bytes:401-500 bytes:501-600 
> > > > 
> > > > no map map for 1-100 map for 401-600 no map 
> > > > 
> > > > 
> > > > 
> > > > 
> > > > 
> > > > 
> > > > 
> > > > 
> > > > 
> > > > Subflow-2: segment-2 segment-4 segment-7 segment-8 
> > > > 
> > > > bytes: 101-200 bytes:301-400 bytes: 601-700 bytes:701-800 
> > > > 
> > > > map for 101-200 map for 301-400 map for 601-800 no map 
> > > > 
> > > > 
> > > > 
> > > > In the above case, the map for data in segment-1 is included in
> > > > segment-3. 
> > > > 
> > > > 
> > > > 
> > > 
> > > 
> > > The question here is why would the stack not put the mapping for segment
> > > 1 on segment 1 itself. And what is the benefit of doing so?
> > > 
> > > 
> > > 
> > > 
> > > 
> > 
> > Well it could just be due to the lack of option space insegment-1. For
> > example, the sender ofsegment-1 at that instant wants to transmit multiple
> > TCP options (e.g.timestamp, SACK, DSS and ADD_ADDR). Obviously, they all
> > cannot fit into optionfield of one segment, and eventually the DSS
> > transmission got slightly delayedby a segment or two.
> 
> The implementation needs to enforce a strict priority of DSS over SACK and ADD_ADDR.
> 

TCP options have evolved over a period of time, and I do not think as such any document/guidelines exist on enforcing priority for one over the other, though it turns out be an interesting topic. Also, more TCP options could come up in future for implementing new features. So, it is likely that implementations would follow different strategy when it come to option priority. 

> 
> 
> 
> If the ADD_ADDR does not fit in the TCP-option space, it can send the
> ADD_ADDR on a pure ACK. The echo-bit in the ADD_ADDR guarantee the reliable
> delivery of it.
> 

I noticed this new feature in RFC 6824-bis to deliver ADD_ADDR in pure ACK and still achieve the reliability. The ability to deliver MPTCP options in pure ACK will be useful, especially for options like ADD_ADDR. 

> 
> 
> 
> Sure, one could argue that favoring SACK over DSS is more important. But I
> think we would need data to justify that. Only very specific traffic
> patterns will fall in this use-case.
> 

The bottomline is that in the event of a map being slightly delayed, i.e., delivered late with respect to the corresponding data, should it result in resetting the subflow? 

As far as the specification is concerned, it could be liberal on accepting such maps, rather than being restrictive. Even if a current implementation cannot support this, future implementations may like to, provided the specification permits this and the implementation is willing to cop up with the associated complexity. 

Best regards,

Anil

> 
> 
> 
> Christoph
> 
> > 
> > 
> > 
> > With regards,
> > 
> > 
> > 
> > Anil
> > 
> > 
> > 
> > > 
> > > 
> > > 
> > > 
> > > 
> > > 
> > > 
> > > Christoph
> > > 
> > > 
> > > > 
> > > > 
> > > > 
> > > > Further, segment-3 cannot combine/cover the data in segment-1 and segment-3 in a "single map", as the data sequence space is not continuous, i.e., some in between data (segment-2) is mapped and transmitted through subflow-2. Here, the map in segment-3 does not even partially cover the data it carries. 
> > > > 
> > > > 
> > > > 
> > > > 
> > > > Both RFC 6824 and the 6824-bis do not restrict the above scenario, and I guess the change proposed now does not permit this to happen. 
> > > > 
> > > > 
> > > > 
> > > > 
> > > > Best regards, 
> > > > 
> > > > 
> > > > 
> > > > 
> > > > Anil 
> > > > 
> > > > 
> > > > 
> > > > > 
> > > > > 
> > > > > 
> > > > > 
> > > > > 
> > > > > 
> > > > > 
> > > > > 
> > > > > That's the reason why we (the MPTCP-upstreaming community) vouch to have this case restricted.
> > > > > 
> > > > > 
> > > > > 
> > > > > Cheers,
> > > > > Christoph
> > > > > 
> > > > > 
> > > > > 
> > > > > 
> > > > > 
> > > > > > 
> > > > > > 
> > > > > > 
> > > > > > 
> > > > > > 
> > > > > > 
> > > > > > Anil 
> > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > Best regards,
> > > > > > > 
> > > > > > > Alan
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > 
> > > > > > 
> > > > > 
> > > > > 
> > > > > 
> > > > > 
> > > > > 
> > > > > 
> > > > 
> > > > 
> > > 
> > > 
> > > 
> > > 
> > >
>