Project

General

Profile

Doc-overview » Random » Filing »

Update (2011-10-18):

(this is the page for the multicast patchset implemented by Linus Luessing for batman-advanced floating around for some time now, which needs and update or replacement by a more up-to-date page)

Current status:

  • we have a working patchset for 2010.2.0 with some limitations (only bidirect connections, no mcast snooping and hence no support for bridged in clients)
  • we have a half-finished implementation which should fix these limitations in T_X's repository

More Info:

  • Batman-adv multicast optimization - how batman-adv optimizes multicast traffic by Linus and Simon [March 2011], slides are attached at the end of this document
  • The specification for the 2010.2.0 version is attached as well

B.A.T.M.A.N.-Adv Multicast Awareness

As batman-adv has full control over all data traffic flowing through the mesh network multicast traffic also falls under its jurisdiction. At the time of writing this document batman-adv handles the multicast traffic by flooding the whole network with it. Although this approach is suitable for common multicast groups sending a small number of packets (e.g. IPv6 neighbor announcements) it fails its purpose when it comes to large multicast data packets (e.g. multimedia streaming). This document aims to provide a concept for multicast optimizations, especially when it comes to these large data packets.

The concept

To be truly multicast aware each batman-adv node needs to perform the following tasks:

  • detection of own multicast groups
  • let the network know which multicast groups this node is belonging to and also learn who else is part of this group
  • deliver the data to the entire mutlicast group in the most efficient way

multicast group sensing

Similar to batman-adv's current HNA sensing mechanisms, batman-adv would look for multicast packets going into its bat0 interface and memorize the according multicast MAC addresses which represent one multicast group id each in IPv6 or a small set of multicast group ids in IPv4.

multicast group participation

Each batman-adv node floods the whole mesh network with the multicast groups it belongs to (it further will be referenced as MCA - multicast sender announcements) and also maintains a list of other nodes in the network belonging to the same multicast groups. In regular intervals unicast "breadcrump" packets are sent to all members of the multicast groups this node is part of unless the other member is not a next hop (these unicast packets just need to mark the forwarding nodes). These "breadcrump" packets are used to inform intermediate nodes which may or may not be part of the multicast group that they are "in the path" of a multicast group.

multicast data packet delivery

Multicast data packets would still be treated as standard batman-adv data broadcast packets, but receive an additional flag to mark them as optimized multicast traffic (the flag is called BAT_MCAST). This packet type will only be rebroadcasted if all of the following conditions are met:
  • the node was marked by a "breadcrump" packet from the same originator for this specific mutlicast group
  • the sequence number of this multicast packet is a new one

This mechanism requires the multicast group infrastructure to be known. Shortly after creating/joining a new group this won't be the case. During the build up phase multicast data packets should still be broadcasted to ensure fast delivery.

New batman-adv packet types

BAT_MCAST

A multicast data packet that shall use an optimized path will still use a batman bcast_packet header, but set packet_type to BAT_MCAST.

mcast_pathsel_packet


/* marks the path for multicast streams /                                                                                                                   
struct mcast_pathsel_packet {                                                                                                                                
        uint8_t  packet_type;  / BAT_MCAST_PSEL /
        uint8_t  version;  / batman version field /                                                                                                        
        uint8_t  srcr6;   / multicast mac address */                                                                                                       
        uint8_t  destr6;                                                                                                                                    
        uint8_t  ttl;                                                                                                                                        
        uint8_t  align;                                                                                                                                      
} attribute((packed));

Pros/Cons

  • As the mcast_path_sel packets are being forwarded via unicast, they'd create a multicast topology with determining shortest, direct paths according to BATMAN's unicast metric. ADAMA-SM for instance is not optimising in this way, instead it tries to reduce the number of forwarding nodes in general (which is nice for the mesh burdon itself, but might create unusable / too long paths for certain multicast receivers in real wireless networks).

Further Extensions

Distribution infrastructure initiation - bursting

To decrease the time needed for the initial infrastructure creation, the multicast announcements could be send with their own packet type instead of being attached to OGMs. They could then be send in a burst of for instance 5 identical packets. Furthermore, these packets could have a rebroadcast-count which means, that also every intermediate node would rebreadcast for instance 5 identical packets again. The wait-time to switch from pure flooding of the multicast packets to the optimised path selection could then for instance be set to 1 instead of the 5 seconds. After that, such designated multicast annuncement packets could be send at a much slower interval as the OGM packets (something like 5-10 seconds for instance).

(Or if there'd be the planned NDP - HELLOs - as well as optional HNA entries (so a missing HNA in an OGM not causing it to be purged on other nodes in the mesh, a designated multicast announcement packet type would not be needed for this. We could just use a small OGM with less information, but still the TQ-value and sequence numbers, too, then as otherwise the ethernet frames' payload would be zero-padded to 46 Bytes anyway which would be a waste.)

still broadcast smaller/rare, but important multicast packets

The smaller the packets, the less harmful the broadcasting would be for the mesh itself. However, the broadcasting would make the transportation of these packtes more robust in most topologies. The critirea for a "small" multicast packet could be:
  • An IPv4 multicast packet from the "Local Network and Internet Work Control Blocks" (224.0.0.0/24, 224.0.1.0/24 - RFC3171). These are for instance IGMP- or mDNS-packets.
  • Well-known IPv6 multicast addresses, having the transient-flag unset. These are for instance the important IPv6 neighbor- and router-discovery packets or mDNS- or DHCPv6-packets.
  • Threshold-triggering: Only if there've been sent for instance 5KB/s during the last second to the same multicast group destination, start building the optimised multicast distribution infrastructure.

For a nice table of multicast IP- and MAC-address ranges, also see this nice wikipedia-article

broadcasting in dense multicast networks

If already about 50% of the nodes are part of the same multicast group, then such an optimised multicast distribution infrastructure's gain by minimising the number of forwarding nodes is not that much and because of the very high maintenance overhead the total-"gain" would even be negative. Therefore, if a multicast member notices that there are about 50% of the nodes in the originator table in the same multicast group, this node would not start sending mcast_pathsel packets and send the multicast data packets via BAT_BCAST instead.

Another optimization for this broadcasting approach in dense multicast networks would be for a node to still check the following:
  • Are all other multicast members I know of behind the same neighbor I just received the multicast data packet from?
    If so, the intermediate node should not rebroadcast this multicast data packet. For this approach multicast packets should never be forwarded as BAT_BCAST packets, a dense/sparse-flag in the batman packet header would be needed instead.

converting BAT_MCAST to unicast if just one member on path left

A node knows, whether there might be a single multicast member of the same group on the forwarding path left (or better: whether all but one multicast members are behind the neighbor we just received the multicast-data packet from) because of the previously received, broadcasted OGMs (+ MCA entries). In this case, the forwarding node can unwrapp the multicast data packet and wrap it into a batman unicast-header to this single destination instead. This will greatly increase the reliability and throughput to such a remote multicast member because the rate selection algorithms being able to select an optimal value instead of just broadcasting it with the default value of 11MBit/s on the one hand and the now acknowledged transfer for the rest of the path on the other.

switching between multicast and unicast forwarding

This is basically an extension of the optimization stated above. But instead of converting a BAT_MCAST to unicast only once at the end of a packets journey, this optimization here also suggests to convert the means of tranportation back and forth: If a multicast data packets journey would get to a crotch, so two different forwarding nodes as next hops for the data, then the packet should be broadcasted by the node on the crotch with the advantage of only blocking the wifi medium once (as the basic algorithm is doing it, too). However, if a node which is part of the distribution infrastructure and knows, that there's just one recipient, just one next hop being a forwarding node, then the multicast data packet should be send as a unicast packet. The packet-type would be a new one (i.e. BAT_MCAST_UNI) and the orig-field needs to be set to the multicast address.

A forwarding node of a certain multicast group distribution infrastructure can detect if it is on a crotch without any additional communication need, it just has to memorise the following more entry: Of one multicast-group's mcast_pathsel stream(s) (the unicast packets maintaining an efficient distribution infrastructure), not only memorize the group and the previous senders mac address, but also the next hop destination mac address(es). If this forwarding node has multiple of these matching next hop destination addresses then it knows it is on a crotch.

So if a forwarding node sitting on a crotch receives a multicast data packet via unicast, then it has to reencapsulate it in a broadcast header with multicast flag. If a forwarding node not sitting on a crotch receives a multicast data packet via broadcast, then it has to reecapsulate it in a unicast header. Otherwise it can forward the packet according to its table without having to reencapsulate it.

NAKs between neighbors

In wireless networks all unicast traffic is being acknowledged and in case of a loss resend until a certain amount of times. We usually don't have this feature for broadcasted packets, especially because of the mobile characteristics of the mesh it can be quite tough to tell on link layer if there was no ACK because of interference or because the neighbor got out of range. It is therefore a lot easier to use NAKs in this case - if a neigbor who is part of the distribution infrastructure detects a missing sequence number, it could request it again and receive it via unicast from the according neighbor. PGM (RFC3208 also using the NAK approach on the transport layer for multicast packets for instance.

Only send MCAs as a receiver(/sender), if there is a sender(/receiver) too

If there is no multicast sender available anyway, then a receiver does not have to announce its multicast member presence because there'd be no need for the distribution infrastructure with no sender anyway. Especially if the multicast sender might not be statically, permanently but adhoc, temporarily available instead, this can reduce the burdon on the mesh network quite a lot if there are also a lot of multicast receivers.

Of course, the other way round, the benefits would be greater if doing it the other way round - receiver-based - if there'd be more multiple senders in the same multicast group and only one receiver there at a time with a very dynamic uptime.

This probably depends on the usage scenarion, but the first option should be the default.

A node can easily detect a receiver-host on its local network by listening to IGMP- or ICMPv6-MDN packets. A sender could be detected by the multicast-destination mac of data packets - however this should not initiate the path maintenance for all kinds of multicast packets as stated above (also IGMP/ICMPv6 are being send via multicast for instance - effectively making any node receiver a sender as well otherwise).

Only build paths from senders to receivers

In most cases, a multicast member won't be a sender and receiver at the same time. In such a case, a selected multicast path from the receiver to the sender is not necessary. Two additional flags MCAST_SENDER and MCAST_RECEIVER could be attached to an MCA. A node receiving an MCA, would then only start the unicasted mcast_pathsel stream if the following requirements are matched:
  • I belong to the same multicast group stated in the received MCA.
  • I am a sender of the MCA's multicast group.
  • The received MCA has the MCAST_RECEIVER flag set.
    Only then an optimised multicast path would be established to the MCA's originator.

Resources: