Speeding up the roaming¶
The base idea consists in exploiting the new forwarding mechanism for data packets described in the Client-announcement page in order to reduce the interruption time a non-mesh client experiences when moving from one mesh node to the next. This procedure of a non-mesh client switching to a new mesh node is called "roaming". At its core it simply is a synchronization issue: Whenever roaming occurs the new mesh node "parent" needs to inform the entire mesh that a particular non-mesh client is now to be found at a new location. Even under ideal conditions it might take a full originator interval in which the client is not reachable at all. The bigger the mesh, the higher the packetloss or originator interval the longer the client has to wait before it can use the network again.
The main idea behind the roaming improvements is to make use of the fact that batman-adv has full control over the payload traffic traveling through the mesh network. In particular, with this mechanism the packet destination is modified along the path, if needed, as the message reaches an up-to-date router. If so data packet is rerouted to the new destination and information in the packet header are updated too (destination and ttvn field) until the mesh network is in sync again. In addition, a roaming advertisement packet is sent to immediately inform the old originator that the client has moved and that it is now located on another node. This message can be seen as an asynchronous update that permits the old node to redirect data packets as soon as they reach it.
Detect a roaming client¶
To detect whether a recently detected non-mesh client is a roaming client or a new client, it is sufficient to search the global translation table for the clients mac address. In case of positive match (the same MAC address is already in the table but pointing to another mesh node) the node has to send a roaming advertisement message to the old mesh node to inform it about the roaming event.
The roaming advertisement message is a unicast TVLV packet and its content is explained in the related TVLV section.
The information contained in this packet is used to update the old node's global table so that, in case of payload packet directed to it while the network is not in sync, with a payload destination equal to 'client address is redirected to 'source address'.
Due to this mechanism, immediately after a roaming the data packets might follow a slightly sub-optimal path (data flow is going through the old node) until the network is in sync again and every node has updated its global translation table. This strategy will help to avoid losing packets that are still flowing to the old node.It could also be the case that before the network get in sync again the client roams once more. Let's assume this scenario: a client X is served by node A and roams to node B, then, before both nodes send the new OGM, client X moves to a third node C. Even in this case the same procedure described above applies, but with some minor extensions due to the second roaming event. In particular this how the nodes involved in the roaming procedure will react:
- X moves from A to B
- B sends a ROAM_ADV to A
- X moves from B to C
- C sends a ROAM_ADV to node A (same mechanism of point 2)
- A updates its global table in order to reroute the packets to the correct (new) destination
- A sends a ROAM_ADV to B to tell it that the roaming phase is over and that the client "came back" (new feature to avoid B to do not re-route packets correctly)
In this way data will be still sent to A first and then rerouted to the correct location with no loss.
Routing data packets¶
The enhanced data routing is built on top of the data routing procedure described in the Client-announcement page. When a mesh node receives a roaming advertisement concerning one of its former non-mesh clients it will mark the client entry with the roaming flag in its own translation table. The mesh node sending the roaming advertisement will do the same. This extra flag is needed as long as the translation tables are not in sync and will have to be deactivated once an OGM with the new ttvn arrives.
When payload data reaches either node (new or old mesh node) they will not only check if the translation table version number (ttvn) of the destination host is outdated but also check if the roaming flag is active. If one of the conditions is met the payload traffic is rerouted if necessary.
Translation table consistency¶
As described in the client announcement document each node computes a set of CRC32 checksum values and floods it through the network using OGMs. The translation table changes triggered by the roaming advertisement message leads to a temporary inconsistent global translation table because the table changes happen outside of the ttvn/CRC32 mechanism.
A node receiving a roaming advertisement message will add a global translation table entry pointing towards the new mesh node, remove the client address from its local table and announces that change with the next OGM. The sender of the roaming advertisement message will complete a similar operation: The client's global entry is removed, the local entry is added and change propagated with the coming OGM. All inconsistencies will be cleaned up by the ttvn/CRC32 system once the new OGMs reach the nodes.
However, neighboring nodes face an interesting problem: If the OGM from the new mesh node containing the "add event" is received before the OGM with "delete event" is coming in, the delete event is silently discarded. But if a node receives the "delete event" first, it has no working route towards the client until the "add event" shows up. To solve this problem the "delete event" comes with a roaming flag (TT_GLOBAL_ROAM) which tells neighboring nodes to use this route until the "add event" arrives. It also excludes this entry from the CRC32 checksum computation.
- MAC conflict: In case of more than one node announcing the same client MAC address, each node will interpret this situation as a "continued roaming" and will start sending roaming advertisement messages multiple times. To reduce the impact of this problematic scenario a roaming-protection has been provided: A client can roam on the same node a fixed amount of times in a fixed length period (look into main.h for the exact values). In case of exceeding the limit, the interested node will not issue any new roaming advertisement unless the protection period is terminated. However, this situation is usually caused by the same client being connected to two mesh nodes (by means of a switch for example) and in this case the Bridge-loop-avoidance will enable nodes to understand that the client is somehow "shared". On the other side, mesh nodes receiving the announcement of the same client from multiple originator, will store a list of node announcing the client and will choose the best (TQ based) to send the data to whenever needed.
A research project has been done on this topic and the final technical report is freely available here: https://eprints.biblio.unitn.it/2269/.