Accurate Time Synchronization for IEEE Based Wireless Networks

Transcription

1 Accurate Time Synchronization for IEEE Based Wireless Networks Jui-Hao Chiang Advised by Prof. Tzi-cker Chiueh Stony Brook University, USA ICNP 2009

2 Outline Motivation IEEE TSF Clock-Jumping Algorithm Experiment Setup and Results Conclusion

3 Motivation Clock Synchronizaiton is important for many control mechanisms in wireless networks Frequency hopping Power management Packet scheduling Consistent ordering for event and sensor value distribution in sensor networks The original IEEE does not perform well in multi-hop wireless networks

4 IEEE Time Synchronization Function (TSF)

5 IEEE TSF Each node keeps a 64-bit micro-second resolution timer counter to implement the Time Synchronization Function (TSF) Beacon packet (frame) are used to carry TSF counter to synchronize with other nodes Different mechanisms in Infrastructure mode Ad hoc mode

6 Infrastructure mode One node is designated as Access Point (AP) AP puts its TSF counter in the beacon frame and sends it out every beacon period (normally 100 msec) All other nodes in the network update its clock based on the TSF counter of AP No big problem; only one commander!

7 Ad hoc mode No AP; each node competes for beacon transmission every beacon period In the beginning of a beacon period Calculate a random backoff delay If the delay timer expires, send out a beacon attached with its local TSF counter Upon receiving a beacon Cancel the pending beacon if any Update its local TSF counter if the frame contains larger counter value (hardware)

8 Ideal condition Beacon Period Fastest node Node 1 Node 2 Transmitted beacon Cancelled beacon Random delay timer expires Time line Three nodes in a single collision domain

9 Problems of ad hoc mode Fastest Node Asynchronism [Huang] Fastest node fails to send out beacon because it loses in the beacon competition Time Partitioning [So] The beacon from fastest node is suppressed and not able to be propagated to all nodes in multi-hop networks

10 Fastest Node Asynchronism Fastest node Beacon Period Node 1 Node 2 Transmitted beacon Cancelled beacon Random delay timer expires Time line Three nodes in a single collision domain; Situation gets worse when more nodes appear

11 Adaptive Time Synchronization Procedure (ATSP) Let faster nodes send out beacon more frequently than slower nodes In each beacon period, if a node receives a larger timestamp, it reduces its beacon transmission frequency Beacon Period Fastest node Node 1 X Node 2 X Time line

12 Approach similar to ATSP [Ye] When a node receives a beacon with larger timestamp Increase random backoff counter range Calculate a lower probability to join the next period beacon transmission Faster nodes will do these in the opposite way Beacon Period Fastest node Node 1 X Node 2 X Time line

13 ABTSF and TATSF [Huang] Adaptive bi-directional TSF (ABTSF) Maintain a list of all node IDs Token concept; with token, the node has higher frequency of sending beacons Periodically handover the token Tiered Adaptive TSF (TATSF) Separate nodes to different tier groups based on the clock speed of the nodes Each group has different beacon frequencies However, the above works still suffer from Time Partitioning problem

14 Time Partitioning Clock island of node A Clock island of node D A B C D Clock rate: D > A > B and C The node with faster clock intuitively has higher priority to send out the beacon

15 Multi-hop Time Synchronization Function (MTSF) [So] Each node maintains a parent variable The node having largest timestamp among all its neighbors Estimating all the received beacon frames Each node schedules its beacon transmission every two beacon periods (200 msec), and staggers with the fastest neighbor If parent node schedules beacon in odd time slot, child node schedules in even slot

16 MTSF (cont.) A B C D D is the parent of C C is the parent of B B is the parent of A First beacon period fastest fastest A B C D (Problem)Expensive steps to build up and maintain all the information Second beacon period

17 Clock-Jumping Algorithm

18 Design Avoid Fastest Node Asynchronism Avoid Time Partitioning Problem Don't want continuous maintenance of explicit tree structure Don't find the fastest node but create one

19 Idea Choose one node to be the fastest node (root) Do not need the fastest hardware chip Leader election problem Every beacon period Root jumps the clock with a significant amount Others do not send beacon spontaneously If a node receives a jumping timestamp Local TSF counter is updated (IEEE virtue) Immediately schedules a beacon transmission with random delay

20 Idea (cont.) root Jump the clock and send root Hmm, a large jump of clock!! (a) time t (b) time t + Adjacent node of root Non-adjacent node of root

21 Beacon Frame Loss Problem Fail, but someone will take care root One beacon is enough! root s (a) Dense network (b) Sparse network Adjacent node of root Non-adjacent node of root

22 Root Failure Problem root Timeout = 2 beacon periods Timeout = 3 beacon periods Each node keeps a timer with a timeout interval proportional to hop count from the root Nearest node discovers first Adjacet node of root Non-adjacet node of root

23 Experiment Setup and Results

24 Implementation Issue Wistron NeWeb a/b/g mini-pci card, Linux , WLAN driver: Madwifi (Atheros) Clock-Jumping Increase one to the the 45th bit of TSF counter inside the WLAN card every beacon period (modifying lower 32-bit causes hardware problem) To prevent from software clock continuity, expose only only lower 44 bits to application (wraps around every 203 days) TSF counter High 32 bit Low 32 bit

25 Monotonicity But if the root node is slower (e.g. the slowest), there might be monitonicity problem Suppose the value of the root nodes TSF counter is X at the time immediately before it is about to jump the clock X[43:0] remains the same after the clock is jumped and is embedded in the beacon frame When a node receives the beacon, our driver expose to application X[43:0]+MCD (maximum clock drift) e.g. if the root node is slowest, X[43:0]+MCD should still be larger than the TSF[43:0] value of the fastest node, and thus guarantees monotonicity Otherwise, the TSF[43:0] of the fastest node will go backwards in the timeline

26 Testbed Setup 4 Computers 25 WLAN interfaces 12- hop linear network Multiple channel multi-hop Two interfaces one Virtual Node Default beacon period 100 msec

27 Find the metrics to measure Use the Pulse computer to send UDP packet every 100 msec All other computers receive it and record the TSF counters of all nodes At any time t (snapshot) GlobalDrift is defined as maximum clock difference between any pair of nodes Among all experiment records MaxDrift is the maximum value of all GlobalDrifts AvgDrift is the average value of all GlobalDrifts Good metrics: MaxDrift?? AvgDrift??

28 Inherent Clock Drift of Hardware 3-hop linear network, no beacon mechanism Zigzag? MaxDrift is not good metric Every 100 msec, the network drifts 1.44 usec Inaccuracy 20 usec!!

29 GlobalDrift of Clock-Jumping Algorithm 3-hop linear network; 6400 msec beacon period Good pattern for synchronization Area below the line = AvgDrift (good metric)

30 Clock-Jumping Algorithm in Multihop Network IEEE is 66% ~ 81% higher than Clock-Jumping Algorithm

31 Fastest Node Asynchronism in IEEE beacon period All nodes are in a single channel Beacon period 6400 msec 10 nodes Fastest node fails to send beacon very often

32 Time Partitioning in IEEE A B C D 3-hop linear network A is clock reference Beacon period 100 msec Clock islands A-B and C-D

33 Impact of Beacon Frame Loss on Clock-Jumping Algorithm The experiments until now are all on a linear topology network without beacon competition 3-hop network, each hop has 6 nodes contending

34 Impact of Root Failures on Clock- Jumping Algorithm 3-hop network, each hop has 6 nodes contending Unplug the power of root node some time after experiment starts One adjacent node becomes the root after 2 beacon periods

35 Conclusion Clock synchronization is an important building block for many wireless protocols, services and applications IEEE suffers from two main problems Fastest Node Asynchronism, Time Partitioning Propose a simple Clock-Jumping Algorithm Some complexity to solve coner cases problem The design, implementation, and evaluation of the proposed algorithm has been shown to improve the time accuracy

36 Thanks Q & A