Re: [multipathtcp] A question related to MPTCP control overhead

[Olivier Bonaventure <[e-mail]>]

Hi Sayee,

> I can think of two ways in which a TCP connection can go down: gracefully or ungracefully. Assume that there are two subflows, SF1 and SF2, and that SF2 is to be brought down. If SF2 goes down ungracefully, we have no cleaner way of informing the receiver that the sender is disabling DSS, and so this leaves us with two choices: (i) continue to use DSS on SF1, or (ii) define a new Option Subtype or a reserved bit of the DSS Option through which the sender can explicitly inform the receiver that it is disabling DSS. I believe that the later choice is the cleaner way of disabling the DSS option. When SF2 goes down ungracefully, we of course need to retransmit the outstanding bytes on SF1, similar to what the MPTCP specification requires.


For a Multipath TCP connection, subflows can fail in other ways as well. 
It is also possible that the source IP address associated to a subflow 
is not assigned anymore to the host, this is the typical example of a 
smartphone that looses WiFi. Another subflow failure scenario is a 
subflow that stops after n retransmissions of the same data. These 
failures do not stop the MPTCP connection.

> The not-so-cleaner way of doing things is fragile and convoluted, so I would not want to discuss that here.
>
> If the connection is closed using the RST bit: after sending RST, the sender immediately closes the socket, so the MPTCP sender has no way of knowing whether the receiver indeed received the RST bit. In fact, the RST bit may be lost in the transit. I would put this case in the category of ungraceful shutdown.

A subflow can fail due to RST sent by a middlebox. This should not close 
the corresponding MPTCP connection.

> If the connection is closed using the FIN bit: Let us represent a SN as x_{i,j} where the subscript i denotes the type of sequence space (SSN or DSN) and the subscript j gives the subflow ID. Let x_{DSN, SF2} be the DSN of the last byte transmitted before FIN is sent on SF2 and that x_{DSN, SF1} be the last DSN mapped to SF1 before the DSS Option is disabled by the sender. In this context we need to consider two possibilities: x_{DSN, SF1} < x_{DSN, SF2} and x_{DSN, SF1} >= x_{DSN, SF2}. Now consider the case x_{DSN, SF1} < x_{DSN, SF2}. When a data byte is received on SF1 that is not covered by DSN we know where to place that data byte with respect to the data byte whose DSN is x_{DSN, SF1} using subflow sequence number of the data byte. Because, eventually, data bytes need to be ordered based on DSN before releasing to the application layer we still need to decide where to place this data byte at the connection-level with respect to x_{DSN, SF2}, and this is where the protocol will run into ambiguity: whether to place the data byte before or after x_{DSN, SF2}. This will lead to protocol dead-lock. Next, we consider the possibility x_{DSN, SF1} >= x_{DSN, SF2}. Here we do not have the ambiguity present in the previous context. The received data byte on SF1 will have to be placed to the right of both x_{DSN, SF1} as well as x_{DSN, SF2} and the actual placement on SF1 will be determined by the SSN of the data byte.
>
> The above analysis helps us to decide how and when DSS Option can be disabled. After sending the FIN on SF2, continue sending the DSS option on SF1 until the largest DSN used on SF1 is at least as large as x_{DSN, SF2} and the state variable NUM_SUBFLOW has transitioned to 1, and then disable the DSS Option. To describe the receiver behaviour, the receiver expects to continue to receive data bytes covered with DS mapping on SF1 until the largest DSN mapped to SF1 is at least as large as x_{DSN, SF2}. After then, starting with the first received byte that is not covered with DSN, the receiver will understand that the sender has disabled DS mapping on SF1.
>
> At this point, I like to believe that defining a new Option Subtype or utilizing an unused bit in the DSS option will be fruitful.
>
> Except in some specific scenarios, like a data center, I don't think a device can know in advance if second interface becomes available during the connection. So if the users goes with TCP then he is at disadvantage because he cannot utilize additional interfaces when they become available. If he opens MPTCP connection hoping that new interfaces may become available in the future, until then the connection has to incur control overhead due to redundant DSS Option.
>
> As a first-cut solution I think we should allow a MPTCP connection not to use the DSS option until it opens the second subflow. To keep things simple, we may say that once DSS Option is enabled it will be enforced until the MPTCP connection is closed.

On smartphones a solution would need to support break-before-make to be 
useful. This means that we cannot terminate the subflow with a FIN 
before deciding to switch to the utilisation of DSS. This fact, coupled 
with middleboxes that can transparently add/remove data from the 
bytestream make the problem difficult to solve


Olivier