Practical Lazy Scheduling in Wireless Sensor Networks. Ramana Rao Kompella and Alex C. Snoeren

Transcription

1 Practical Lazy Scheduling in Wireless Sensor Networks Ramana Rao Kompella and Alex C. Snoeren

2 Distributed Rate Adaptation Problem: In wireless networks (e.g., sensor nets, ) radios consume significant power Opportunity: Where radios have multiple transmit rates, reducing rate may reduce energy consumption Challenge: Find a distributed, online transmit scheduling algorithm

3 Radio Power Consumption Transmit Electronics 50nJ/bit Transmit Amplifier 100pJ/bit/m 2 d [HeinzelmanISLPED2000] Receive Electronics 50nJ/bit Transmit Amplifier most dominant energy consumer except for short distances (5-10m) Electronic power dictated by Moore s Law, is decreasing with improvements in technology Transmit RF power more or less remains the same

4 Key Idea: Exploit Power/rate Tradeoff Many radios have multiple transmit power/rate levels Shannon s laws suggest that energy in joules/bit is convex in rate of transmission So, why not transmit all packets at the lowest speed? Energy required Rate of transmission Problem: Delay increases if packets sent too slowly

5 Transmit Scheduling Algorithm Power level for a given rate Different radios have different energy-delay curves Henceforth, only rate selection Application layers Scheduling Algorithm MAC Layer Wireless medium Radio with multiple Tx rate and power levels

6 Prior Work In Reducing Tx RF power Mechanism : Shurgers et al. show how to adjust rate discretely for QAM radios Show energy gains for a single sender Study energy-delay tradeoff Single Sender algorithm: Uysal,et al. analyze theoretical scheduling algorithm. Limitations: Single sender with knowledge of arrivals Offline algorithm Continuous Radios

7 Our contributions Discretization: Offline algorithm for radios with discrete set of Tx rates Single Sender : Online algorithm for power conservation Multiple Senders : Distributed self-tuning algorithm for multi-node case Implementation over CSMA/CA

8 Talk Outline Single Sender Offline Algorithm (Uysal et.al) Online Algorithm for single-sender case Multiple Senders Distributed online algorithm for multi-node scenario Implementation over CSMA/CA Conclusions

9 Optimal Offline Algorithm [uysal01] Assumptions: All arrivals known in advance No time deadlines for individual packets Radios are continuous All packets need to be transmitted within the interval with no additional delay

10 Optimal Offline Algorithm Transforms a known set of arrivals into a schedule of transmission durations Schedule takes advantage of packet inter-arrival times Schedules packets such that output rate approximates input rate of packets

11 Optimal Offline Algorithm 100 Inter-arrival Time Packet Number

12 Optimal Offline Algorithm 100 Transmission Duration Distributes the packets as uniformly as possible Packet Number

13 Online Algorithm Approach Estimate the amount of traffic Choose transmission rate based on this estimate Two phases in pipelined fashion Look-ahead : Buffer packets Scheduling : Schedule packets with optimal bit-rate

14 Simple example From applications Wireless medium buffer scheduler

15 Simulation Description Poisson Bursty traffic arrival patterns in paper (qualitatively similar) Variable bit rates from 100Kbps to 1Mbps Results for different bit rates in paper (qualitatively similar) Fixed packet size of 10Kbits Two Key Metrics Average Power = total Energy / total Time Average Delay = total Delay / total packets Energy function used (Uysal et.al) w(t) = 10 6 /0.06 t ( 2 (0.12/t) - 1)

16 Results : Power savings (single node) Average Power in nw 3.5e+08 3e e+08 2e e+08 1e+08 Offline Online Lookahead=500ms Online Lookahead=1000ms Online Lookahead=2000ms Increasing lookahead 15% higher at 0.8 8% higher at 0.5 5e+07 5% higher at Offered Load / Maximum Capacity

17 Results : Average Delay (single node) Average Delay ms Offline Online Lookahead=500ms Online Lookahead=1000ms Online Lookahead=2000ms Delay ~ 1/2 lookahead Increasing lookahead Offered Load

18 Performance summary (single-node case) Online algorithm achieves power savings competitively (within 5-8% for light loads) with offline algorithm Average delay suffered by packets bounded and predictable Increasing the look-ahead achieves better power savings but also introduces higher average delay

19 Talk Outline Single Sender Offline Algorithm (Uysal et.al) Online Algorithm for single-sender case Multiple Senders Distributed online algorithm for multi-node scenario Implementation of L-CSMA/CA Conclusions

20 Multi-node Scenario Transmit Range of Yellow Node Shared broadcast channel One owner manages every node (no malicious nodes, global power savings) CSMA/CA based MAC protocol Time synchronization (usec) Non negligible packets to send

21 CSMA/CA example Load = 4 Node A Node A senses the medium and if idle transmits packet Again, Node A senses the medium, finds idle If collision due to no ack, backoff random time Shared Wireless channel Load = 3 Node B Node B senses the medium and if idle transmits packet

22 Goals Power Saving : Achieve power conservation using transmit rate adaptation Self-tuning: No prior information about arrival patterns, number of nodes, etc. Easy Deployment : Apply no changes to any of the existing layers Global properties: Global fairness and throughput, delay are important

23 Simple extension to multiple-node case Quantize time into intervals All nodes follow two phase pipelined protocol Each node transmits packets in scheduling phase based on look-ahead Throughput = total packets/ total time decreases

24 Total Load Estimation Need a distributed way to communicate the overall load in the scheduling phase Naïve way would be to indicate before the start of every interval the load on the network Do not know apriori how many nodes have packets to send Allocating a fixed amount of time to perform load discovery in every interval is wasteful

25 Distributed Online Algorithm First node that gets access sends its packets slowly Nodes with packets snoop to determine the Tx rate and the sender If node is sending the first time, each node increases its estimate Nodes account for CSMA/CA delays

26 Multi-node Example Estimate =4 Node A Load = 4 Packet 1 transmitted at rate s.t. transmission Duration = 10/4=2.5ms Packet 2 transmitted at rate ( )/5 =1.25ms (Estimate Converged) 1. lookahead=10ms 2. all packet sizes same 3. collision backoff times are subtracted Load = 3 Estimate = 6 Node B Node snoops the channel transmission And transmits its packet at (10-2.5)/6=1.25ms

27 L-CSMA/CA Architecture Network Layer Buffer Scheduler Data Link Layer CSMA/CA MAC Physical Layer Snoop information about sender and rate of transmission

28 Four nodes : Power vs Load for different Lookahead values 7e+08 6e+08 CSMA/CA Poisson L-CSMA/CA LA=1000 Poisson L-CSMA/CA LA=2000 Poisson L-CSMA/CA LA=4000 Poisson Average Power 5e+08 4e+08 3e+08 2e+08 1e+08 99% power savings Increasing lookahead Graceful degradation to CSMA/CA with decreased lookahead 93% power savings Offered Load

29 Throughput vs Number of Nodes L-CSMA/CA Poisson CSMA/CA Poisson Observed Throughput % Number of nodes throughput loss up to about 10% with about 100 nodes

30 Fixed Lookahead : Average Delay vs Number of nodes L-CSMA/CA Poisson Packet Size = 1Kb L-CSMA/CA Poisson Packet Size = 10Kb Ganeriwal et.al, dynamically modified the level of output buffering used in DMS to tradeoff power savings against packet delay Average Delay Number of nodes Seems to grow Linearly with number of nodes Smaller packet size allows early convergence

31 Conclusions We use energy-rate tradeoff in real radios with multiple transmit bitrates Our online algorithm achieves power saving within 5% of the offline for the single sender case In multiple node scenario, we achieve up to 99% power savings for four nodes in 20% load as compared to regular CSMA/CA Demonstrated that our algorithm is self-tuning, and requires no changes to existing layers

32 Scheduled Power And RaTe Adaptation