Google Summer of Code 2011 - Ideas¶
Categories + Usecases to improve¶
- Increasing Robustness
- Improving Mobility / Convergence Speed
- Improving Indoor performance
- Improving VoIP group multicast performance
- Improving nomadic scenarios (circular + chain topology)
Dynamic OGM+NDP intervals¶
Brief description: A batman-adv node shall select originator interval rates according to the dense and dynamics in its closer environment.
Depending on the usage scenario, people can adjust the bandwidth being used for batman-adv's route finding algorithms. Usually people were advised to increase the originator interval if the mesh network is small but needs fast route refresh rates or to decrease it if the mesh network is mostly a static setup with a lot of nodes. It would be great if a batman-adv node could determine the dynamics of the mesh network it is currently participating in on its own so that this option would not have to be administrated from a person anymore. For instance in combination with the 'B.A.T.M.A.N. protocol overhead reduction' project, the local broadcasts could be automatically increased if there are not that many direct neighbors.
OGM/NDP Hidden Node Problem Avoidance¶
Brief description: Implement a suitable solution against hidden node broadcast packet problem.
Especially in indoor scenarios where corners and thick walls are involved, hidden nodes can be a severe problem. Usually activating RTS/CTS is a common way of solving this problem at least to some degree. However, RTS/CTS can only be applied for unicast packets. Therefore a node sending a lot of data packets to another one, even with RTS/CTS those packets will interfere with BATMAN's broadcast packets (e.g. NDP packets or OGMs). The effect is, that the transmit quality of a node sending NDP packets which does not see the data packet transmission will greatly decrease.
A suitable solution shall be implemented. This might be based on time slots, multiple interfaces or channel switching and might be implemented within batman-adv or an extra module. See here for a more detailed description of the problem and a solution proposal.
OGM/NDP Congestion Avoidance¶
Brief description: Implement a congestion avoidance mechanism for batman-adv's routing control packets against congestion caused by high data transfer rates.
In case of a mesh network congested due to e.g. a lot of UDP packets, the packet loss rate of batman's OGMs or NDP packets increases severely, as neither OGMs/NDP nor UDP have a congestion control mechanism. For one thing, lost OGMs lead to a high degradation of batman's convergence speed. For another, lost NDP packets lead to route flipping; in the worst case some of these chosen routes are not usable. batman-adv has the advantage of being able to control the whole packet, both data and control packets, within the mesh network. Therefore time spaces for OGMs/NDP packets could be allocated, as already described in "OGM/NDP Hidden Node Problem Avoidance" / "Or prioritized queues could be introduced.
B.A.T.M.A.N. Routing Framework¶
Brief description: Separate B.A.T.M.A.N. routing algorithm from routing infrastructure in batman-adv.
Many features in batman-adv are routing algorithm independent: Link layer routing, link layer fragmentation, multicast optimizations, NDP, interface bonding. To allow other routing algorithms to use this already well tested infrastructure, the B.A.T.M.A.N. routing algorithm shall be separated more clearly from the routing infrastructure. Ultimately, other routing protocols shall be able to control the kernel space mesh routing infrastructure from userspace.
Improve Throughput Bonding¶
Brief description: Improve performance gain of the throughput bonding mode.
The current bonding mode implementation simply sends a data packet in a round robin fashion to the interfaces available for bonding. However, the throughput gain seems to be only about 66% instead of doubling the throughput in case of TCP, even if there is no other interference involved. This seems to be the case due to a heavy reordering issue.
Furthermore, the interface with the lowest capacity is a bottleneck for the throughput bonding performance: In case of one interface of 1MBit/s throughput and five more with 2MBit/s throughput, the actual throughput will be just 6x 1MBit/s. However in mixed wireless and wired networks, it might be desirable to accumulate the throughput of a Gigabit-Ethernet interface and a 802.11g interface.
Therefore checks for links' capacities (detect full queues etc.) to gain performance of IF1 + IF2 + ... instead of min(IF1, IF2, ...) shall be implemented. If an interface is busy, it shall be skipped in this round robin cycle. Furthermore simple per hop pre-ordering on batman-adv's layer shall be implemented to increase the TCP performance.
Unit Test Infrastructure¶
Brief description: Create a flexible unit test infrastructure to increase batman-adv's stability.
As batman-adv resides in the Linux kernel any crash in batman-adv has severe consequences for the whole system. For instance a deadlock or aceesing invalid memory in batman-adv will freeze the whole system. Typically most embedded routers have a watchdog to restart the device automatically in such situations. However in critical usage scenarios, a reboot of the device might not be tolerable. Therefore a unit test infrastructure shall be created which for one thing tests and verifies single functions and functionalities within the batman-adv code base. And for another it shall stress test the complete software for different virtualized, static and dynamic topologies.
IPv6 Stateless Autoconf Gateway Solution¶
Brief description: Implement a gateway solution for IPv6 stateless autoconfiguration
batman-adv operates on layer 2, therefore intelligent layer 3 gateway handling is not straightforward. batman-adv currently supports intelligent gateway handling for DHCP (both IPv4 and IPv6), in that it forwards any DHCP discovery or request packet to the best (according to batman's metric) gateway available only, via unicast. However, this is not doable for IPv6 stateless autoconfiguration, as usually not the client is not actively requesting an IP address, but the router is advertising it's gateway capabilities periodically. However, such a stateless configuration is usually more desirable than a stateful configuration as in DHCP in dynamic mesh networks as it allows smoother transitions between subnets and is not as prone to packet loss. Therefore a solution for IPv6 Stateless autoconfiguration shall be implemented. This may be done by exploiting the router preference fields in the router advertisement messages, RFC4191 , and/or by limiting the range of router advertisements with the help of batman-adv's broadcast packet's TTL field or by introducing a new TQ limit field. A TQ limit field could be dynamically adapted to minimze the overhead of such router advertisements in case of many gateways but ensuring the availability of the best/closest of all gateways in the mesh network.
Improve Broadcast Data Performance¶
Brief description: Optimize the flooding of broadcast data to reduce overhead and increase its packet delivery ratio.
So far broadcast data packets are flooded through the whole mesh network. As broadcast packets do not have an ARQ mechanism to reduce packet losses like unicast data transfer has, batman-adv (re)transmits broadcast packets on each hop by default. In dense topologies this can introduce unnecessary overhead, in sparse topologies the packet delivery ratio might be too low. Therefore a smarter mechanism than classic flooding shall be implemented to improve batman-adv's performance for broadcast data packets. This may done taking link qualities in the local neighborhood into account and adjusting the number of rebroadcasts dynamically. Or by introducing MPR-like mechanisms.
Reduce ARP (/ND) latency and overhead¶
Brief description: Reduce ARP latency and overhead by implementing an ARP cache.
batman-adv operates on the link layer, therefore any network protocol (e.g. IPv4/6) needs to perform an MAC address lookup for any IP address through the mesh network first. As ARP packets are broadcast packets, batman-adv simply floods them through the whole mesh network. This can create quite some overhead in large scale mesh networks or high latencies in case of mesh networks with poor links. Therefore an ARP cache shall be implemented on each node so that they could answer ARP requests directly to any host on their segment instead of flooding it through the whole mesh again. Something similiar may be implemented for IPv6's Neighbor Discovery (ND) mechanism.
Dead node fast path switching/invalidating¶
Brief description: Improve batman-adv to respond quicker to node/link failures and avoid packet drops in case of link failures.
When a node notices the breakdown of a neighbor (see routing scenarios to get an idea about the conditions), this node could send any data packet, which it would usually send over this neighbor to either its second best hop if available (which does not always have to be the case due to OGM forwarding policies). Or it could send the packet back to the next hop towards the source again. With the help of sequence numbers, any node on the 'backtracking' path (the backtracking path can be different from the usual path in case of asymmetric links) could notice that a path became invalid very quickly.
Link layer FEC/ARQ/Fragmentation module¶
Brief description: Create a kernel module that can be added onto an interface which performs link layer optimizations (FEC, ARQ, Fragmentation, ...).
The ARQ and FEC mechanisms provided in 802.11 for unicast data packets is mainly optimized for one hop scenarios only. While the packet delivery ratio on one hop might be fine, it might not be usable for multiple hops in a mesh network. Therefore a link-layer kernel module shall be implemented which performs Forward Error Correction and Automatic Repeat reQuest with dynamic parametrizations, to achieve a certain, configured target packet delivery ratio. Furthermore, the fragmentation from batman-adv could be moved to this link layer module, too. Such a module would then provide a virtual interface which batman-adv would use instead of the actual wifi/ethernet interfaces.
- Link quality / Packet delivery ratio measurement improvements
-> weighted/exponential moving average
-> testing / performance measurements of / improving NDP, find and improve "performance bottlenecks"
- Multicast Optimization Algorithm enhancements
-> implement reactive tracker packet mode
-> decrease latency of reactive tracker packet mode: attaching small data packets to tracker packet, if MTU fits (otherwise broadcast) in react. mode? ...?
-> ultimately, make proactive tracker mode obsolete
-> optimize number of rebroadcasts (we have the info about the number of next hop nodes)
-> don't send tracker packets if <= 1 destination or > 50% of all nodes are destinations
-> Enhance HNA, host network announcements, improvement scheme to also support MCAs, multicast announcements
- Further HNA roaming improvements
- Network Coding?
- Packet signing: Add a signing mechanism to identify OGM's sender and drop malicious nodes
- built-in bandwidth test tool ?
- Enhance live link quality monitoring: bisect -> dot-files -> graphs -> to (live) video? (+ adding horst tool information, wifi interface stats?)
- Add further info to batman-adv vis servers/clients for debugging networks
- bisect, include initial state
- live vis in map (HTML5 + Openstreetmap?)
- multiple interface support in Mesh 3D
- ap51flash GUI; ap51flash multi-flash on single interface
- Android/Maemo/Meego porting + maintenance!
- Cooperative work with Pidgin guys? (video/audio/file-transfer/bonjour in pidgin and improving its + batman-adv's combined performance in a mesh network?)