Considerations for MPTCP operation in 5G

One of the goals of 5G systems, as outlined in , is to enable low latency services and access to local data networks where mobility anchors can be deployed close to devices, thereby satisfying use cases with stringent transmission delay and high reliability. Mobility in 4G networks, as described at the architecture level in , was based on a central mobility solution that made it difficult to relocate mobility anchors closer to the end user. In contrast, 5G uses a distributed mobility solution based on multiple anchors providing different IP addresses as the device moves from one area to another. The base scenario in this section is: a 5G device connected to a single-homed server is in an area with no usable Wifi coverage. An application using MPTCP sends traffic over a single subflow, over the cellular air interface. Then, as the device moves, the 5G device reacts to mobility events. Additionally, we briefly discuss cases where the device switches from a Wifi to a cellular backup connection, and other cases where both MPTCP peers are 5G mobile devices. In 5G, every unit of a network connectivity service (PDU session) has a type which can be IP (IPv4 or IPv6), Ethernet or unstructured. While session continuity is supported for all types, we will focus on PDU sessions of IP type primarily. Different PDU sessions will typically correspond to distinct network interfaces on the device (though this is not explicit in the standard, and some implementations may possibly behave differently). In 4G networks, session continuity is enabled by anchoring a PDN Connection (as PDU Sessions are referred to in 4G networks) to a P-GW which allocates an IP address to the mobile device: PDN connection and IP address allocation are maintained as long as the device remains attached to the network, even when the device moves around. In 5G, different types of session continuity can be provided, and are indicated by a "Session and Service Continuity" (SSC) mode value of 1, 2 or 3 (defined in section 5.6.9). Every PDU session is associated with a single SSC mode, which cannot be changed on this PDU session. The following sub-sections will study how 5G handles each SSC mode, and potential effects on MPTCP.

IP Address Session Continuity Service Type The "session continuity service type" (SCS type) of an IP address allocated by the network can be used as input by MPTCP implementations. It has been defined for on-demand mobility management , as: FIXED IP address: valid for a very long time, for session continuity and IP address reachability. SESSION_LASTING IP address: valid for the lifetime of an IP session, even if the mobility host moves. NON_PERSISTENT IP address: which does not provide session continuity nor IP address reachability. GRACEFUL_REPLACEMENT IP address: similar to a non-persistent address, but adding a limited graceful period for the transition from one address to another. This information needs to be conveyed to the device by the network that allocates the address: for example, as described in , the SCS type of IP address may be conveyed through router advertisements. The MPTCP stack may obtain the SCS type of a local IP address using a programmatic API. This information does not directly tell which SSC mode the application is using, nevertheless it is sufficient for MPTCP to appropriately and non-ambiguously decide how to handle new IP addresses in the context of SSC in 5G. Notification of a New IP Address from Session Continuity As in non-5G cases, MPTCP starts handling the initial socket created by the application. The initial connection attempt triggers the creation of the first PDU session (or the reuse of an existing one), associated with a network interface and an IP address on this interface. Later on, the network decides to use a new anchor, e.g., when a mobility event occurs or when a Local Data Network becomes available. The network then triggers the modification of the existing PDU session, or creation of a new PDU session, and ultimately a new IP address is provided to the 5G device, respectively on the same or a new network interface. The 5G stack on the device is aware of the relationship between the old and new IP addresses, and can therefore notify the MPTCP stack. Such a notification may include both the old and new IP addresses , which ensures that the MPTCP stack has all necessary information (for example, the MPTCP stack may then obtain the SCS type of the new IP address, and the validity lifetime associated with the old IP address, if it is still up).

In SSC mode 1 the same network anchor is kept regardless of device location. An application running on the device will therefore be able to keep using the same IP address on the same interface. Additionally, in SSC mode 1, the network may decide to add and remove, dynamically, additional network anchors (and therefore IP addresses) to the PDU session, while always keeping the initial one. PROPOSED MPTCP BEHAVIOR SPECIFICATION: These steps are applicable when the initial IP address returned by the network stack is FIXED or SESSION_LASTING. MPTCP should not close all subflows originated from this original IP address at any point during the session, since this IP address is the only one that is guaranteed, under normal circumstances, to be maintained over time for this application. At any time during the session, a new IP address of SCS type NON_PERSISTENT may become available. MPTCP may create new subflows for the application, using this IP address (this IP address is likely to provide shorter path subflows, but may disappear at any time).

SSC mode 2 has a break-before-make behavior. When the device leaves the service area of its first network anchor, the network stops using it and starts using a new second network anchor closer to the device. (Such service areas may have a highly variable size depending on network deployments.) On the device, this can result in the currently used network interface being brought down, and after a short time a new network interface being brought up. The time between these 2 events is not standardized and implementation dependent. Break-before-make within cellular technology When MPTCP is active on cellular only, this break-before-make behavior is similar to the existing break-before-make capability usually used in cellular/Wifi offload (section 3.1.3 of and section 2.2 of ). A similar MPTCP behavior is therefore needed: wait for a given time for a new IP address to be configured. As per , to benefit from this MPTCP resilience feature, the application should not request using a specific network interface. Cellular and Wifi Additionally, when Wifi is active and cellular is used as backup, MPTCP implementations should also support this break-before-make behavior to maintain a usable backup IP address on cellular. In rare cases where a switch-to-cellular backup would be needed during a break-before-make transition on cellular, MPTCP's existing break-before-make capability can ensure MPTCP waits for a new cellular-facing IP address to be available. PROPOSED MPTCP BEHAVIOR SPECIFICATION: These steps are applicable when the initial IP address returned by the network stack is NON_PERSISTENT. At any point in time, the current NON_PERSISTENT IP address may be taken down by the network stack. The MPTCP stack should wait for another NON_PERSISTENT IP address to be made available by the network stack. If such an address is made available within a given time limit, the MPTCP stack should create new subflows using this new address (effectively following the existing break-before-make behavior present in MPTCP). Additionally, if an initial backup IP address is a NON_PERSISTENT address, the MPTCP stack should consider any subsequent NON_PERSISTENT IP address as a backup IP address in replacement of the initial NON_PERSISTENT address.

SSC mode 3 has a make-before-break behavior. When the device leaves the service area of its first network anchor, the network selects a second network anchor closer to the device, and either creates a new PDU session (i.e. new IP address on new network interface) or share the existing PDU session (i.e. new IP address on same network interface). The first network anchor keeps being used for a given time period, which is communicated to the device by the network using the "valid lifetime" field of a prefix information option in a router advertisement (, ). 5G does not mandate a specific range for this valid lifetime. The first/older IP address should not be used to create any new traffic ( section 5.5.4). In some implementations, the network (SMF) may decide to release the first network anchor as soon as it stops carrying traffic. There is no limit set by the 5G standard for the number of concurrently used network anchors. We expect that in usual cases the first network anchor will be released before a third network anchor starts being used. Nevertheless, to our knowledge nothing prevents a 5G system deployment to allow a third network anchor to be selected while the first one is still in use. On the 5G device, when using SSC mode 3, mobility will therefore result in a new IP address being configured, either on the same network interface initially used, or on a different interface. In general, an application will see a single cellular-facing IP address, and during transient phase it will see 2 IP addresses (with a possibility for more than 2 concurrent IP addresses on some 5G system implementations). In cases where the server is single-homed and the Wifi interface is down, and assuming a full-mesh path manager policy, there will be in general one subflow, and from time to time, temporarily 2 subflows (or more on some 5G systems). In cases where two mobile 5G devices are communicating with each other over MPTCP and with the same assumptions on Wifi and path manager policy, there will be in general one subflow, and from time to time, temporarily 2 or even more rarely 4 subflows (again, possibly more on some 5G systems). MPTCP must create new subflows when a new IP address on a same or a new cellular-facing network interface becomes available to the application. MPTCP may keep using the first subflow during a transient phase. Here are some considerations related to this transient phase: When compared with simply waiting for the first IP address to be brought down, ramping down usage of the first subflow will not incur inefficiencies from resending lost segments. This may especially help low-latency applications by avoiding throughput drop. Assuming a lowest-rtt-first scheduling policy is used, after the initial TCP slow start, the shortest path subflow should typically carry the most traffic. Ramping down should ideally start after the initial slow start is over. To make sure the ramping down completes before the interface is brought down by the network, the MPTCP stack should be aware of how long will the first network anchor be kept in use, e.g. through configuration or communication with the local 5G stack. Ramping down and closing flows on the first network anchor as soon as possible will help recycling network resources more rapidly. This is especially true in cases where more than 2 network anchors may be used concurrently. There may be some level of contention between subflows during the transient phase, since they share the same air interface, and especially if they share the same PDU session and QoS marking. The shortest path subflow may therefore not reach its full capacity during the transient phase. Additionally, the shortest subflow must not be closed during the transient phase (even if it is less efficient for some reason), to avoid losing all connectivity at the end of the transient phase. To avoid this issue, the MPTCP stack could for example follow a policy not to close any subflow created using the latest IP address, during the transient period (in SSC mode 3). In cases where cellular is used for backup, there is a possibility that the switch to using backup occurs during a transient phase. To support this case, MPTCP should keep creating and releasing subflows as described above, even when cellular subflows are used as backup, to ensure that the backup is always usable. When a backup event occurs during a transient phase, MPTCP should use the subflows associated with the most recent cellular-facing IP address, i.e. corresponding to the latest/closest network anchor. PROPOSED MPTCP BEHAVIOR SPECIFICATION: These steps are applicable when the initial IP address returned by the network stack is GRACEFUL_REPLACEMENT. At any point in time, a new GRACEFUL_REPLACEMENT IP address may be made available by the network stack. The MPTCP stack must create new subflows using this new address, gracefully transfer traffic to these new subflow(s), and close subflow(s) using the previous GRACEFUL_REPLACEMENT IP address before its scheduled closing (known by obtaining the valid lifetime of the IP address from the operating system). Additionally, if an initial backup IP address is a GRACEFUL_REPLACEMENT address, the MPTCP stack should consider any subsequent GRACEFUL_REPLACEMENT IP address as the new backup IP address, in replacement of the first GRACEFUL_REPLACEMENT IP address.

The SCS type of an IP address is only available locally, which limits to only one peer the MPTCP behavior specified in sections , and . This can potentially cause problems: A remote may for example close subflows that should be maintained e.g. closing all subflows to a SESSION_LASTING IP address or to the latest GRACEFUL_REPLACEMENT address. A remote may also not behave optimally in some cases, e.g. because it cannot distinguish between a transition from GRACEFUL_REPLACEMENT to GRACEFUL_REPLACEMENT, from a temporary addition of a NON_PERSISTENT address to a SESSION_LASTING address. There are two possible ways to handle this: either explicitly transmit the SCS type over MPTCP signalling, or use heuristics on the remote peer (e.g. never close a subflow created by a peer, mirror behavior of peer, etc.). In the explicit solution, a MPTCP peer provides the SCS type of IP addresses to its remote peer, using existing or new options (mp_capable, add_addr, mp_join, or an experimental option). When compared to using heuristics, transmitting the SCS type can lead to a less complex and more explicit implementation, which can help better handling unusual or error cases, and adapt to future evolution of 5G or other mobility management systems.