The Importance of Being Opportunistic

Transcription

1 High Performance Switching and Routing Telecom Center Workshop: Sept 4, The Importance of Being Opportunistic Sachin Katti Dina Katabi, Wenjun Hu, Hariharan Rahul, and Muriel Medard

2 Bandwidth is scarce in wireless Can we send more while consuming less bandwidth?

3 Current Approach Relay Alice A lice s p acket B o b s p acket Requires 4 transmissions Can we do it in 3 transmissions? Bob B o b s p acket A lice s p acket

4 A Network Coding Approach Relay Alice A lice s p acket B o b s p acket Bob B o b s p acket A lice s p acket 3 transmissions instead of 4 Save bandwidth

5 Network Coding Routers mix bits in packets, potentially from different flows Theoretically shown to achieve capacity for multicast No concrete results for unicast case

6 How to apply network coding? State-of-the Art Multicast Given Sender & Receivers R S R Given Flow Rate & Capacities R Min-Cost Flow Optimization Find the routing, which dictates the encoding

7 How to apply network coding? State-of-the Art Multicast Given Sender & Receivers Given flow rate & capacities In Practice Unicast Many Unknown Changing Sender & Receivers Unknown and bursty flow rate Min-Cost Flow Optimization Find the routing, which dictates the encoding?

8 How to apply network coding? State-of-the Art Multicast Given Sender & Receivers Given flow rate & capacities In Practice Unicast Many Unknown Changing Sender & Receivers Unknown and bursty flow rate Min-Cost Flow Optimization Find the routing, which dictates the encoding Opportunism

9 Opportunism (1) Opportunistic Listening: o Every node listens to all packets o It stores all heard packets for a limited time

10 Opportunism (1) Opportunistic Listening: o Every node listens to all packets o It stores all heard packets for a limited time A D B

11 Opportunism (1) Opportunistic Listening: o Every node listens to all packets o It stores all heard packets for a limited time o Node sends Reception Reports to tell its neighbors what packets it heard Reports are annotations to packets If no packets to send, periodically send reports

12 Opportunism (2) Opportunistic Coding: o Each node uses only local information o Use your favorite routing protocol o To send packet p to neighbor A, XOR p with packets already known to A Thus, A can decode o But how to benefit multiple neighbors from a single transmission?

13 Efficient Coding A C D B Arrows show next-hop

14 Efficient Coding Bad Coding C will get RED pkt but A can t get B LU E pkt A C D B Arrows show next-hop

15 Efficient Coding Better Coding Both A & C get a packet A C D B Arrows show next-hop

16 Efficient Coding Best Coding A, B, and C, each gets a packet A C D B To Arrows XOR n packets, show next-hop each next-hop should have the n-1 packets encoded with the packet it wants

17 But how does a node know what packets a neighbor has? Reception Reports But reception reports may get lost or arrive too late Use Guessing o If I receive a packet I assume all nodes closer to sender have received it

18 Putting It Together A C D B

19 Putting It Together A C D B

20 Putting It Together A C D B D on t reorder packets in a flow Keeps TCP happy No scheduling No packet is delayed

21 Beyond Fixed Routes B heard S s transm ission S S transmits A B No need for A transmission Route Chosen by Routing Protocol D Opportunistic Routing [BM05] Piggyback on reception report to learn whether next-hop has the packet cancel unnecessary transmissions

22 Opportunism Unicast Flows arrive and leave at any time No knowledge of rate No assumption of smooth traffic O A B N C F K M D G E

23 Performance

24 Emulation Environment We use Emstar o R eal code sim ulator o radios o Power Level: 200mW o 11Mbps bit rate o Simulated radio channel

25 Recall Our Simple Experiment XOR Relay Alice A lice s p acket B o b s p acket Bob B o b s p acket A lice s p acket 3 transmissions instead of 4 25% throughput increase

26 1 Throughput (KB/s) No Coding 330 KB/s XOR 270 KB/s 0 Practical artifacts make network coding perform poorly Network coding requires broadcast But broadcast has no backoff more collisions

27 Ideally, design a back-off scheme for broadcast channels In practice, we want a solution that works with off-the-shelf drivers/cards Our Solution: Pseudo Broadcast Piggyback on unicast which has synchronous Acks and backoff o Each XOR-ed packet is sent to the MAC address of one of the intended receivers

28 1 XOR with Pseudo-Broadcast Improves Throughput Throughput (KB/s) No Coding 330 KB/s XOR with Pseudo-Broadcast 495 KB/s XOR 270 KB/s Improvement is more than 25% because MAC gives nodes equal bandwidth shares Without coding, relay needs twice as much bandwidth With coding, all nodes need equal bandwidth

29 Larger experiment 100 nodes 800mx800m Senders and receivers are chosen randomly Metric: Total Throughput of the Network

30 Opportunistic Coding vs. Current Network Throughput (KB/s) Opp. Coding Our Scheme Opp. Coding (no guessing) No Coding No Coding A Unicast Network Coding scheme that works well in No. realistic of flows situations

31 Preliminary Implementation Results Linux Kernel MAC Click Elements (part of Roofnet) Only Opportunistic Coding & Pseudo Broadcast (No opportunistic Listening) Relay Alice Bob

32 Ratio of throughput with coding to without coding Implementation Results Network Coding Doubles the Throughput Bit Rates (Mb/s)

33 Conclusion First implementation of network coding in a wireless network Learned Lessons o Be opportunistic (greed is good!) o Can do a good job with multiple unicast o 5x higher throughput in congested networks o Preliminary implementation results show throughput doubles

34 The Wireless Environment Multi-hop wireless networks (e.g., Roofnet)

35 Opportunistic Coding vs. Current Network Throughput (KB/s) Our Scheme No Coding 0 A Unicast Network Coding scheme that works well in No. realistic of flows situations