Network Coding: technical documentation

When forwarding packets in a wireless mesh network, airtime can be saved by exploiting the information already present at the destinations of a forward. One way of doing this is Network Coding, where two packets are XOR'ed together before transmission. To extract one of the included packets, the destination must simply XOR the network coded packet with the other included packet:

  (p1 XOR p2) XOR p2 = p1

For more information about the concept, read the NetworkCoding introduction.

Implementation Details

The implementation is based on 5 steps:

  1. Detect neighbors that are in range of each other: This enables the relay to determine if a neighbor is likely to overhear packets from another neighbor.
  2. Hold back unicast packets before forwarding: To get two packets to code, one must be buffered until the next arrives.
  3. Code and transmit combined packets: This step searches the buffer with packets to be forwarded and checks if the destinations will be able to decode.
  4. Save overheard and transmitted packets: To decode a a packet, the receiver must know one of the two combined packets, which can either originate from the receiver itself or be sniffed from another neighbor.
  5. Receive and decode network coded packets: When receiving a coded packet, the overheard and saved packets are searched for the packet needed to decode.

Each step is explained in detail in the following sections.

1. Detect neighbors in range of each other

A relay node exploits the TTL field of OGMs to detect when two neighbors are in range of each other:

Node A emits an OGM, which is received by Node C and Node R.

Node C decreases the TTL retransmits the OGM from Node A. The OGM is again received by Node R, who compare the TTL of the two received OGMs.

If the difference if only 1, Node R can conclude that Node C is in range of Node A.

2. Hold back unicast packets before forwarding

Whenever a unicast packet is ready to be forwarded, it is passed to the network coding system, where it added to a buffer (only if not coded with an already existing packet). The buffer is organized as a hash table with entries of src-dst pairs. Each entry has a linked list with packets that all come from a specific previous hop and are directed to a specific next hop (which is decided by the routing part of batman-adv). The linked list works as a FIFO:

A worker function traverses the list periodically and transmits timed out packets. The timeout is by default 10ms.

3. Code and transmit combined packets

Before adding packets to the forward buffer, the buffer is searched for opportunities to network code. The algorithm to find such opportunities takes its starting point in the packet currently being added to the buffer, say /p1/:

  1. The destination of a potentially network coded packet must be in range of the source of /p1/, so we loop nodes in range of the source of /p1/.
  2. Since we also need the destination of /p1/ to be in range of another node, we loop the nodes that the destination is in range of.
  3. For each pair of nodes in the two loops, we check of buffered packets and network code if any.

4. Save overheard and transmitted packets

To be able to decode network coded packets, a node must possess one part of the coded packet. This is achieved by buffering both packets originating at the node and packets overheard by the node. The buffer is structured in the same way as the forward buffer in step 2, which makes is fast to search for decoding packets. To overhear packets from other nodes, the wireless interface must be put in promiscuous mode.

5. Receive and decode network coded packets

To transmit a network coded packet to two destinations, the address of the additional receiver is put the batman-adv header, and thus this receiver is required to use promiscuous mode. Furthermore, the header contains information needed to identify the id of the original packets contained in the network coded packet.

To identify the packet needed to decode, the receiver uses the source of the network coded packet, and a field in the header with the address of the sender of the original packet. These addresses are used to create a src-dst pair, which is used to locate the correct entry in the hash table, in which the packet list is searched for a matching CRC, which is also given in the header of the network coded packet.

When the needed packet is located, its data is XOR'ed with the network coded packet, and the original unicast header is restored so that it can be processed as usual by batman-adv.

Packet Format

The header of a coded packet is rather large, as it has to carry the information needed to identify the packet needed to decode and restore the headers of the two decoded unicast packets:

  0                   1                   2                   3
  0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
 | Packet Type   |    Version    |   First TTL   |  First TTVN   |
 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
 |                         First Source                          |
 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
 |         First Source          |    First Orig Destination     |
 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
 |                    First Orig Destination                     |
 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
 |                          First CRC32                          |
 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
 |  Second TTL   | Second TTVN   |      Second Destionation      |
 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
 |                       Second Destination                      |
 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
 |                         Second Source                         |
 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
 |        Second Source          |    Second Orig Destination    |
 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
 |                    Second Orig Destination                    |
 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
 |                          Second CRC32                         |
 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
 |        Coded Length           |          Payload ...          |
 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
Explanations of each field:
  • header: The usual header required for all batman-adv packets. It has the TTL of the first of the two combined packets.
  • first_ttvn: Version number of the used translation table for the first combined packet.
  • first_source: Address of the node who sent the first of the two combined packets to the relay.
  • first_orig_dest: Address of the originator that the first of the two combined packets should to routed to.
  • first_crc: CRC checksum of the first of the two combined packets.
  • second_ttl: TTL of the second combined packets.
  • second_ttvn: TTVN of the second combined packets.
  • second_dest: Address of the second of the two nodes that should receive this packet.
  • second_source: Address of the node who sent the second of the two combined packets to the relay.
  • second_orig_dest: Address of the originator that the second of the two combined packets should to routed to.
  • second_crc: CRC checksum of the second of the two combined packets.
  • coded_len: Length of the shortest of the two combined packets.

A field "first_dest" (address of the first receiver of the network coded packet) is not needed because it is already present in the mac header following the network coding header.

The header can be logically divided into two main parts: One part for each of the combined packets. For each packet, the header specifies the destination (although the first destination is given in the MAC header), the previous source, the final originator destination and a CRC32 checksum of the packet.

To identify the packet needed to decode the network coded packet, the receiver must reconstruct a unique identifier, consisting of source, destination and packet checksum, from the received network coded packet. The first field (source) is the original source, which is given as either first_source of second_source in the header. The second field (destination) is the source of the relay, which is found as the source field in the MAC header. The third field (checksum) is found as either first_crc or second_crc in the header of the network coded packet.

After identifying the right packet and decoding, the receiver reconstructs the original unicast header with the fields from the network coded header. E.g. mac_dest, mac_src, ttl, ttvn, orig_dest.

The field coded_len is needed when the packets of different length are network coded, as only a part of the resulting packet will actually be coded. The coded_len field tells the receiver the length of the shortest combined packet, as the receiver of this only knows the length of the longer packet.

Todo

There are few ideas that could make the implementation of network coding perform even better:
  • Support setups with multiple interfaces. (Need to make sure that nc_nodes are on the same interface.)
  • Configure interfaces to promiscuous mode automatically.
  • Hold back packets only if the network is congested.
  • Determine allowed holding times based on the tick time of the system.
  • Support fragmented unicast packets.
  • Construct a better algorithm to select which receiver to put in the MAC header destination field (weighted TQ).

Limitations

Although network coding can provide a significant gain in throughput, there are a few limitations, which should be known:
  • To see a gain, two flows must intersect at a relay, who can network code packets.
  • Relays hold back packets for up to 10 ms to wait for other packets to be coded.

ogm_step1.png (6 kB) Martin Hundebøll, 11/08/2012 10:40 pm

ogm_step2.png (4.6 kB) Martin Hundebøll, 11/08/2012 10:40 pm

packet_pool.png (13.8 kB) Martin Hundebøll, 11/08/2012 10:48 pm