Opportunistic Routing in Multi-hop Wireless Networks* A. Zubow Sanjit Biswas and Robert Morris, SIGCOMM 2005, pdos.csail.mit.edu/papers/roofnet:exor-sigcomm05/roofnet_exor-sigcomm05.pdf

Traditional wireless routing Abstraction wireless radio looks like a link wireless network looks like a graph Routing protocol computes shortest paths on graph Routing protocols: AODV, DSR, TORA, DSDV, OLSR Link metrics: hop count, ETX, ETT, WCETT packet packet A B src dst packet C 2

Radios aren’t wires Every packet is broadcast src dst 1 1 1 1 2 2 2 2 3 3 3 3 4 4 4 4 5 5 5 5 6 6 6 6 C Every packet is broadcast Reception is probabilistic 3

Basic Idea In wireless networks, use of channel diversity => better performance e.g., time diversity, spatial diversity, frequency diversity Probabilistic reception of a single broadcast at different receivers provides such a form of diversity. 4

Key idea Delayed binding of next hop Node broadcasts packet No next hop specified Due to probabilistic reception, different receiver nodes may/may not receive packet Receiver “closest” to destination becomes the next hop, i.e., it will forward the packet towards destination. 5

Key idea A B src dst C packet packet packet packet packet packet 6

Why this might work … src N1 N2 N3 N4 N5 dst 75% 50% 25% Assumes probability falls off gradually with distance Best traditional route over 50% hops: 3(1/0.5) = 6 tx Throughput  1/# transmissions ExOR exploits lucky long receptions: 4 transmissions 7

Why this might work … Assume independent losses 25% 100% N2 25% 100% src dst 25% 100% N3 25% 100% N4 Assume independent losses Traditional routing: 1/0.25 + 1 = 5 tx ExOR: 1/(1 – (1 – 0.25)4) + 1 = 2.5 transmissions 8

Challenges in protocol design Agreement protocol: nodes must agree on the subset of them that received the packet Scheduling protocol: to minimize collisions between different nodes by avoiding simultaneous transmission Efficiency in forwarding: node “closest” to destination that receives packet should become forwarder Only nodes that are really useful should participate 9

ExOR batching Send batches of packets for efficiency rx: 53 rx: 85 rx: 99 rx: 40 rx: 22 rx: 23 rx: 57 rx: 88 rx: 0 N2 N4 tx: 0 ` tx:  12 tx: 100 tx: 34 src dst N1 N3 tx: 31 tx: 23 Send batches of packets for efficiency Node closest to the dst sends first Other nodes listen, send remaining packets in turn Repeat schedule until dst has whole batch 10

Protocol Design: Outline Source sends packets in batches Source includes a priority list of forwarders, ordered by “distance” to destination in every packet header. Scheduling: Lower priority nodes wait for higher priority nodes before transmitting. A “batch map” is used for agreement: indicates the highest priority node that has received each packet included in the packet header updated from higher priority nodes back towards lower priority nodes provides an acknowledgement 11

Protocol Design Layer 2.5 solution Batching: BatchSz BatchID PktNum Scheduling: FragNum FragSz Agreement: Batch Map 12

Priority ordering of forwarders src dst N1 N3 Goal: nodes “closest” to the destination send first Sort by (modified) ETX metric to dst Nodes periodically flood ETX “link state” measurements Path ETX is weighted shortest path (Dijkstra’s algorithm) Source sorts and includes list in ExOR header 13

Digression: ETX primer Measurement of link ETX metric: Measure forward probability of loss pf using periodic HELLO packets Then, ETX = 1/pf If the link is i.i.d Bernoulli with loss probability pf, then ETX = expected number of retransmissions for success. Path ETX = sum of link ETX along path Can be computed using Djikstra or Bellman Ford 14

Digression: ETX primer 15

Batch-map: Reliable summaries tx: {2, 4, 10 ... 97, 98} summary: {1,2,6, ... 97, 98, 99} N2 N4 src dst N1 N3 tx: {1, 6, 7 ... 91, 96, 99} summary: {1, 6, 7 ... 91, 96, 99} Repeat summaries in every data packet in a batch map Cumulative: what all previous nodes rx’d This is a gossip mechanism for summaries Propagate batch map back from destination to sender 16

Scheduling Goal: only one node sends at a time Src sends entire batch Dst sends 10 empty packets with batch maps Then highest priority node starts transmitting whatever it has received Other nodes: on receiving pkt, update batch map and estimate sending rate of current sender use FragSz and FragNum to estimate sender’s time to completion Assume (in absence of information) that each higher priority node transmits at least 5 packets Schedule repeats after lowest priority node transmits. 17

Scheduling 18

Some more protocol details … Towards the end of a batch, overhead of scheduling and agreement spread out over fewer packets Send last 10% of packets of a batch using traditional routing Too many forwarders = high overhead Source runs a ExOR simulation at the beginning and only chooses those nodes which transmitted over 10% of packets 19

Using ExOR with TCP Web Server Client PC TCP ExOR Batches (not TCP) Node Gateway Proxy Web Proxy ExOR Batching requires more packets than typical TCP window 20

Evaluation Details 1 kilometer 21

Evaluation Details Traditional Routing ExOR Click router user-space implementation Directly tx and rx raw Ethernet frames: libpcap like interface IEEE 802.11b, Intersil Prism 2.5 chipset, 200mW transmit power, pseudo-IBSS 65 Node pairs 1.0MByte file transfer 1 Mbps 802.11 bit rate 1 KByte packets Link loss measurements used to compute ETX metric offline Traditional Routing ExOR 802.11 unicast with link-level retransmissions Hop-by-hop batching UDP, sending as MAC allows 802.11 broadcasts 100 packet batch size 22

25 Highest throughput pairs 3 Traditional Hops 2.3x More opportunities for making progress 2 Traditional Hops 1.7x 1 Traditional Hop 1.14x Batch maps ACKs are more robust and efficient 1000 ExOR 800 Traditional Routing 600 Throughput (Kbits/sec) 400 200 Node Pair 23

25 Lowest throughput pairs 1000 ExOR 4 Traditional Hops 3.3x 800 Traditional Routing 600 Throughput (Kbits/sec) 400 200 Node Pair Longer Routes ExOR can use longer asymmetric links 24

Throughput improves by 200%!! 25

Packets that fall short help! 26

Packets that travel further help! 27

ExOR moves packets farther 58% of Traditional Routing transmissions 0.6 ExOR Traditional Routing Fraction of Transmissions 0.2 25% of ExOR transmissions 0.1 100 200 300 400 500 600 700 800 900 1000 Distance (meters) ExOR average: 422 meters/transmission Traditional Routing average: 205 meters/tx 28

Discussion Late binding of next hop is similar to selection diversity RTS to multiple potential receivers Decide “best one” after receiving CTS Another example of diversity use is “opportunistic scheduling” Different users have different channels Send to the user with the best channel Cellular downlink scheduling, OAR 29

Discussion Implicit design assumptions: Carrier sense is good enough RTS/CTS handshake does not buy us much more Spatial reuse gain is negligible scheduling protocol ensures only one node to transmit at a time All of these assumptions are satisfied when mesh networks are small and don’t have too many high hop-count paths 30

Discussion Scheduling in ExOR ensures that higher priority nodes transmit earlier than lower priority nodes prevents collisions between transmissions of the same flow No inter-flow collision avoidance is done I don’t think there are easy ways to handle the scheduling problem in this case 31

Discussion Scheduling may lead to collisions priority inversion 32

Discussion For long routes, unicast comparison is unfair hop by hop unicast transmissions does not do any spatial reuse In IEEE 802.11 settings, there will be some throughput boost due to spatial reuse No multiple flow evaluations are carried out. 33

Conclusion ExOR uses a form of selection diversity to bind the next hop only after receiving the current packet Radio is not a wire: treat it as a broadcast medium instead ExOR gives 2x throughput improvement ExOR implemented on Roofnet 34