surveys detailing the protocols and their advantages and disadvantages, but we feel that they don't give a general im- age of TDMA techniques, which would be ... Wireless sensor networks (WSNs) have proven to be use- ful in many fields and ...
The Second International Conference on Sensor Technologies and Applications
TDMA Protocol Requirements for Wireless Sensor Networks Victor Cionca Electronic and Computer Engineering University of Limerick Ireland
Thomas Newe Electronic and Computer Engineering University of Limerick Ireland
Abstract
memory, and most important, they are battery-powered so they have very limited power supply. Despite this, sensors are expected to run unattended for months, even years; considering that the most power consuming task is radio communication, there is a lot of research in designing energyefficient medium access control (MAC) protocols [11]. The main schemes for MAC protocols are CSMA/CA (carrier sense multiple access collision avoidance) and TDMA (time division), as FDMA (frequency division) requires complex hardware and CDMA (code division) has high computational demands. In the remainder of this paper we explore TDMA techniques. We will also present the setbacks of CSMA protocols and compare CSMA to TDMA. TDMA protocols create a schedule for network activity: each node is assigned at least one slot in a time frame, which is considered to be the number of slots required to get a packet from each source to the sink. The protocol can have a global network frame with the same length in all nodes; this makes synchronization easier but requires the maximum slot number to be propagated throughout the network, which introduces more traffic. Other choices are neighborhood and per node time frames, which reduce the level of traffic but are more difficult to implement and require more complex mechanisms for synchronization. Slots are assigned to nodes for sending or receiving data, and when there is no activity the node can shut down it’s radio interface, thus saving power. It is very important that the slots are assigned in a way that prevents transmission interference, or conflicts. These are of two types:
TDMA is probably the best choice for an energy efficient multiple access protocol for wireless sensor networks. As such there exist many approaches of increased complexity, using different algorithms and combining information from other layers or multiple access techniques. There are many surveys detailing the protocols and their advantages and disadvantages, but we feel that they don’t give a general image of TDMA techniques, which would be useful in choosing the protocol for a network deployment, or even for someone trying to develop a new MAC protocol. We analyze the different algorithms used in TDMA protocols and we present a list of characteristics that a robust TDMA protocol requires. We also give a brief survey of several protocols.
1
Introduction
Wireless sensor networks (WSNs) have proven to be useful in many fields and environments, from hospitals [18] , to battlefields [14] and wildlife habitats [13]. They make life easier for scientists and the common man and provide an easier interaction between man, machine and environment. Research in WSNs is fueled by the differences between sensor networks and traditional networks. Mainly, while traditional networks are human-centric, sensor networks are tools, functionality-based and data-centric. Sensors gather data and send it to the base station periodically, upon receiving a request (demand-driven), or immediately after data has been sensed (event-driven). This creates traffic shapes throughout the network that are different from traditional: in periodical or demand-driven mode data the network will be inactive most of the time, but in certain periods all the nodes will start sending or forwarding data that will flow from the sensors towards the base-station. This is the main traffic type in WSNs and is known as converge-cast. The other difference is in the hardware capabilities of the sensor nodes (motes). They have limited processing power,
978-0-7695-3330-8/08 $25.00 © 2008 IEEE DOI 10.1109/SENSORCOMM.2008.69
Vasile D˘adˆarlat Facultatea de Calculatoare Universitatea Tehnic˘a Cluj-Napoca Romˆania
• Primary conflicts caused by two adjacent nodes sending data at the same time. • Secondary conflicts caused by two non-adjacent nodes sending data to a a single, third receiver. This is known as the “hidden node” problem and is handled in CSMA/CA with the RTS/CTS messages. Two or more nodes that don’t conflict can be active in the same slot. In a multi-hop topology, this provides spatial
30
power consumption. Also, TDMA achieves better channel utilization as the number of nodes is higher, which matches the WSN profile. In table 1 we give a qualitative comparison of the two schemes, while for quantitative results we recommend the technical report by Warrier and Rhee [20]. There are WSN applications where TDMA is not appropriate and topology, mobility or expected data rate should be analyzed before selecting the MAC protocol. TDMA doesn’t adapt well to network changes which are quite frequent in sensor networks due to node failure. Also, nodes have to be synchronized such that the clock drift between nodes be negligible compared to the slot size. Synchronization problems can lead to high latency and power consumption caused by interference from overlapping schedules. While CSMA protocols may be suited for event-driven WSN applications with dynamic topologies, TDMA protocols work best in applications with periodical or on-demand data gathering. Monitoring with non-critical data (so no real-time demands requiring urgent response) is a good example, where sensors gather data for a long time, then they wake up and start sending it to a central node or gateway [7]. We start by looking at the general algorithms used in TDMA scheduling while in section 3 we examine some of the novel approaches. In section 4 we discuss the influence of network topology over algorithm choice and in section 5 we make several suggestions to the behavior of an ideal TDMA protocol. In the last section we give a brief description of protocols used in sensor networks
Table 1. Comparison of TDMA and CSMA/CA TDMA CSMA/CA Power consumption Low High Bandwidth utilization Maximum Low Preferred traffic level High Low Dynamic (network change) Poor Good Effect of packet failure Latency Low Synchronization Crucial reuse. Also of great importance in TDMA is choosing when to update the schedule. The nature of the sensor nodes can lead to frequent topology changes due to node failures, node mobility or new nodes being added to the network. Rescheduling can either be done on demand upon topology change, or periodically, where we distinguish between the following: • scheduling at the beginning of a session, where a session assumes that all source packages have arrived at the sink • recomputing at the beginning of each time frame • the schedule is computed just at network establishment, which is not recommended
1.1
A comparison of TDMA and CSMA in the context of WSN
The classic CSMA/CA protocol is the IEEE 802.11 standard [1] based on RTS/CTS and DCF (distributed contention function) mechanisms. It prevents collisions by repeatedly probing the medium and waiting for it to become available. However, channel utilization decreases when the number of nodes increases, as bandwidth is wasted by taking repeated back-off to avoid collisions. In the context of WSNs with on-demand and periodical data gathering these characteristics of CSMA protocols are not efficient. Constant probing of the medium in the long inactive periods can quickly drain nodes’ batteries. Also, when the network is active, considering the large number of nodes and high density in WSNs, the CSMA protocols are again considered inefficient due to reduced channel utilization. Woo and Culler in [21] have proven the inappropriateness of 802.11 for WSNs, from the point of view of energy management. To adapt CSMA to WSNs techniques like low-power listening or overall network duty cycle scheduling, as implemented in S-MAC [23], are used. TDMA protocols will only schedule the activity of the network and in that period, all nodes will be active. In the idle times between data gathering sessions nodes can turn off the radio interface and lie in a sleep state. Thus, the main and most important advantage of TDMA over CSMA is low
2
General TDMA Scheduling
Depending on how the schedule is created, TDMA protocols are centralized or distributed. Most scheduling algorithms are derived from the map coloring or graph coloring problem and their correctness can be proven mathematically, but there are some which take a collaborative approach, similar to human or process communication. These are harder to prove but sometimes give better results.
2.1
Centralized Algorithms
The straightforward approach for TDMA scheduling is centralized algorithms: • the base station gathers information about the network topology • network is represented as graph, taking conflicts into account • color the graph • compute schedule based on the fact that nodes with the same color send in the same slot
31
An example for a centralized algorithm is RAND [16]. RAND is a graph coloring algorithm; information about the network topology is mapped to a graph. The graph edges are first put in random order and then they are assigned colors so that any edge will have a different color than it’s preceding neighbors. Centralized algorithms are very fast and simple, but they require a lot of traffic to build full topology information at the base station. They are best suited for static star topologies with a low number of networks, but not for multi-hop WSNs with many nodes, especially if the topology changes often.
and confront it with their two hop neighbors, which leads to a global, evolving schedule. These protocols first have a schedule exchange and set-up phase where nodes will access the medium in CSMA/CA mode. As presented in Figure 1, the first one to gain access will establish it’s schedule based on the amount of data it must send. Nodes keep a table of all advertised schedules and will try to set-up their own based on this and on local, two-hop topology information: if two nodes don’t interfere, they can use the same slots. The result is not a global schedule of fixed length, but rather neighborhood schedules which can be easily modified. We also refer to these algorithms as first unoccupied slot algorithms. An example is DRAND (Distributed RAND) [4], which maps a solution of the dining philosophers problem on the TDMA one. Nodes can be either philosophers (transmitters) or forks (receivers), where philosophers must contend to win access to forks. They use priorities and lotteries to establish the winner which will first enquire about the fork’s availability, that is, if the fork hasn’t already been acquired by another node. Therefore, each transmitter keeps track of his receivers and vice-versa. The problem that arises in distributed schedules is their adaptability to topology changes caused by node mobility, failure, or adding more nodes. The distributed graph coloring can be improved if an incremental update of the schedule is used, as stated in [22]. The algorithm will assign new nodes a color that avoids conflicts in the two-hop neighborhood. If such color doesn’t exist, a new one will be created, but since this adds to the size of the time frame, periodical maintenance procedures are required, where the entire schedule is recomputed. The same approach could be used for collaborative algorithms but without the periodical maintenance; therefore the adaptability of collaborative algorithms is limited to the number of new nodes that can join a neighborhood and keep a short schedule.
2.2
3
Figure 1. Schedule establishment in collaborative algorithms. Nodes 1a and 1b choose at the same time, as they don’t conflict.
• the schedules can be refined if information about the number of packets at each node is available • the base station distributes individual schedules to nodes
Distributed Algorithms
Novel approaches to TDMA for WSNs
In the past section we have seen the advantages and limitations of traditional TDMA protocols. Designing the TDMA protocol that suits sensor networks best requires more complex techniques, which we explore in this section.
In sensor networks distributed TDMA algorithms are preferred over centralized ones as they are more suited to large numbers of nodes and they do not need full topology data. There are approaches to distributed graph coloring like PEDAMACS [5] and [3]. In [5] the protocol uses a token passing algorithm, where the base station will broadcast a token which stores already assigned colors. The token must return to the source and each node will have taken note of surrounding colors. Distributed graph coloring algorithms are very computational and so require nodes with a higher level of resources than we generally find available in the sensor nodes. Collaborative algorithms are distributed algorithms suitable for WSNs because of reduced complexity and communication. The nodes each construct their own schedules
3.1
Using Routing Information
Considering that the main traffic used in wireless sensor networks is convergecast, the order in which the nodes are active in the TDMA frame can influence network performance. If the network traffic has a direction, as in routing paths, scheduling the nodes in random order can introduce high latency: a node scheduled to be active before his predecessors on the path will have all the path waiting one time frame. The worst case scenario is where the schedule order
32
nounce receivers of the amount of data they have to send. When receivers have received all the data they can turn the radio interface off and save more energy. Also they will not be contending for their neighbors slots which will reduce the back-off windows in the contention area. Figure 2. Trivial case for scheduling with routing information.
4 The Impact of Network Topology on Algorithm Choice When we discuss about types of TDMA schedule, it is important to differentiate between the topologies for which the protocols are implemented. Sensor networks will have clustered, tree and also mesh topology, and the behaviour of TDMA is different in every case. Clustered topologies are basically star topologies and the scheduling problem is a simple one: in order to avoid conflicts, all the nodes must have a separate slot so it is just a matter of selecting the desired order in which nodes will send to the cluster head. In cluster topologies TDMA approaches try to solve the problem of asymmetric energy consumption where nodes will take turns as cluster heads. Mesh or multi-hop topologies require the most complex TDMA schedules. Usually graph coloring algorithms are used and the focus is on providing spatial reuse and finding the shortest length time frame that fits the network. However, pure mesh topologies are useful if data is randomly exchanged between nodes, as in an Ad-Hoc network. In a sensor network, the main traffic being converge-cast, the topology is reduced to a tree rooted at the base station and with source nodes as leaves. It must be noted that nodes on different routing paths (so not neighbors in routing tree) may be neighbors in the real network and suffer from interference and collisions. Using an algorithm for mesh scheduling in a sensor network can be regarded as inappropriate. Because data is forwarded towards the base station there is an implicit ordering for the send-receive-sleep cycle of nodes that should be respected. Therefore, for sensor networks with a mesh topology we recommend implementing a TDMA scheme which uses information obtained from the routing algorithm. According to [6] the behavior of a wireless sensor network changes dramatically in real world situations with only around 150 deployed nodes. Among the main factors are asymmetric and/or long links, which can take up to 15% of the network links. Schedules that don’t take these into account risk collisions or latencies or even failing slots, where a node should send a packet to its peer but cannot because they are connected by an asymmetric link. Therefore it is very important to design a TDMA protocol with these irregularities in mind. For graph coloring algorithms, a good solution, though costly, is directed conflict graphs. Distributed collaborative algorithms are more exposed to asymmetric and long links; first the irregularities
is the opposite of the routing path order, that is, the sink is first and source is last. In this case sending one packet will take n time frames if there are n nodes on the path. Solving the right order problem is best achieved if routing information is used in the scheduling. To pass data from node to node without delay one node must receive while his predecessor transmits, creating a sort of daisy chain. The trivial case of the straight line is presented in Figure 2. There are however the issues of communication nodes, where two or more paths converge in one node that is not the sink. Then the node must sequentially receive packets from it’s children. Slot assignment joint with routing information has been done in Lin [10] by creating single path, virtual circuits which achieve minimum latency; in Arisha et al [2] in a clustered, centralised way where the cluster head assigns slots in the routing direction; and finally in DMAC [12] where transmit/receive slots on the routing path are assigned consecutively so that nodes do not wait for data. We will examine the particularities of this last protocol later on.
3.2
Hybrid TDMA and CSMA/CA
The last type of algorithms we have come across are the hybrid algorithms which mix CSMA and TDMA. They address the fact that TDMA is only efficient on high levels of network traffic where there is a constant stream of data exchange. When nodes in the schedule have no more data to send slots are unused and delays are introduced. This is opposite of CSMA which achieves maximum efficiency when there is little traffic. Protocols like ZMAC [17] achieve dynamical adaptation of the MAC to traffic load by using a CSMA phase at the beginning of every TDMA slot. The back-off period is shorter for the slot owners, but if they don’t use the slot any contenders can gain access to it if they have data to send. This way all the slots in the schedule will be used. The disadvantage is that nodes will spend a part of the TDMA slots in listening which adds to the power consumption. An improvement to hybrid protocols has been proposed in [19] to prevent nodes from listening when the sender has no more data to transmit. They introduce a prescheduling phase after slot assignment, when transmitters will an-
33
Z-MAC or Zebra-MAC [17] is a hybrid CSMA/CA and TDMA protocol which employs explicit contention notification (ECN) messages to adapt the behavior to traffic level in the two-hop neighborhood. Nodes will contend for a slot as long as its owner doesn’t experience interference or packet loss. The protocol also uses a local synchronization scheme. Because the protocol uses low-power listening, it is more CSMA than TDMA; the slots are not strict even under heavy contention. Z-MAC can use the prescheduling scheme [19] for data direction awareness. Joint TDMA and routing [2] uses routing information in TDMA. The protocol is centralized and slots are assigned in the routing tree using depth-first or breadth-first node visiting techniques. Test results and our analysis show good performance energy-wise, with low traffic overhead. However the protocol does not employ spatial reuse and does not solve secondary conflicts. The protocol will introduce great latencies in large networks. DMAC [12] is a TDMA protocol designed for data gathering in sensor networks. It is different from the others at it doesn’t use traditional TDMA frame and slot assignment. The nodes will schedule their activity based on the number of children and the amount of upstream data. Conflicts are solved using more data messages or data prediction. DMAC thus behaves as a distributed hybrid protocol which uses routing information.
have to be fully detected in the neighbor discovery phase. Then asymmetric peers need to compute their schedule in the right order, so that the unconnected node precedes the connected one. Otherwise the peers may be active in the same slot, thus causing collisions. The preferred solution to long and asymmetric links should be hybrid protocols, as they handle collisions in the CSMA phase and eliminate the effects of network irregularities.
5
Requirements from a TDMA Protocol
We have studied many TDMA protocols and scheduling algorithms. We suggest that when implementing a TDMA protocol, the following should be respected: • Deal with secondary conflicts, which are harder to prevent than primary ones. • If the topology is multi-hop, provide spatial reuse. • Use routing information to eliminate delay caused by wrong node order in the schedule. • Use a distributed algorithm to minimize communication overhead. • Deal with node synchronization to prevent unplanned collisions from happening.
7
• Take into account practical issues like asymmetrical and long links, which can cause collisions and are hard to trace.
6
Conclusions
TDMA scheduling by itsself may not make the most appropriate MAC protocols for sensor networks, due to low tolerance to topology changes. However, we can see that more and more MAC protocols are using some of the TDMA techniques presented here to improve the performances of traditional CSMA/CA protocols. The amount of power saved when using a TDMA protocol is undeniable and is of great importance considering the requirements of sensor networks. Although the requirements we presented are straightforward, they are difficult to meet in practice, and our protocol comparison table has showed that. We feel that there is still work to be done in finding the right TDMA protocol for sensor networks and solving problems like conflict resolution and tolerance to topology changes.
TDMA in WSNs
In WSN literature there are many excellent surveys of MAC protocols, like Kredo II et al [8] or Langendoen [9]. Therefore, we only give a brief review, to support our affirmations. In Table 2 we present their compliance with the list of characteristics we have given for an ideal TDMA protocol. PEDAMACS [5] or Power Efficient and Delay Aware MAC for Sensors employs centralized as well as distributed graph node coloring to find the shortest schedule possible and minimize latency. The protocol constructs a conflict graph from the original network to represent conflicts. The authors give formal proofs of the protocol’s correctness and also give test results. Distributed edge coloring [3] is a distributed graph edge coloring algorithm, based on the Misra and Gries algorithm [15]. The authors use an augmented graph to solve the edge activation problem discussed before, and also prove the correctness of the protocol. A very complex protocol, it doesn’t adapt well to topology changes.
Acknowledgment The authors wish to thank the following for their financial support: • The Embark Initiative and Intel, who fund this research through the Irish Research Council for Science, Engineering and Technology (IRCSET) postgraduate Research Scholarship Scheme.
34
PEDAMACS Distributed coloring Z-MAC Joint with routing DMAC
Centralized Distributed both distributed
Table 2. Comparison of TDMA protocols Hybrid Routing Secondary Dynamic Info Conflicts no no good no no no good no
distributed
yes
centralized distributed
Synchronization
Topology
none none
tree mesh mesh
tree
good
no
no
if use prescheduling yes
no
no
yes, neighborhood yes
yes
yes
has issues
yes
not needed
References
tree
[13] A. Mainwaring, D. Culler, J. Polastre, R. Szewczyk, and J. Anderson. Wireless sensor networks for habitat monitoring. In WSNA ’02: Proceedings of the 1st ACM international workshop on Wireless sensor networks and applications, pages 88–97, New York, NY, USA, 2002. ACM. [14] W. M. Merrill, F. Newberg, K. Sohrabi, W. Kaiser, and G. Pottie. Collaborative networking requirements for unattended ground sensor systems. Aerospace Conference, 2003. Proceedings. 2003 IEEE, 5:5–2153, 8-15, 2003. [15] J. Misra and D. Gries. A constructive proof of vizing’s theorem. Inf. Proc. Lett., 41:131–133, 1992. [16] S. Ramanathan. A unified framework and algorithms for (t/f/c)dma channel assignment in wireless networks. In Proc of IEEE INFOCOM, pages 900–907, 1997. [17] I. Rhee, A. Warrier, M. Aia, and J. Min. Z-mac: a hybrid mac for wireless sensor networks. In SenSys ’05: Proceedings of the 3rd international conference on Embedded networked sensor systems, pages 90–101, New York, NY, USA, 2005. ACM. [18] L. Schwiebert, S. K. S. Gupta, and J. Weinmann. Research challenges in wireless networks of biomedical sensors. In MobiCom ’01: Proceedings of the 7th annual international conference on Mobile computing and networking, pages 151–165, New York, NY, USA, 2001. ACM. [19] W. Wang, H. Wang, D. Peng, and H. Sharif. An energy efficient pre-schedule scheme for hybrid csma/tdma mac in wireless sensor networks. Communication systems, 2006. ICCS 2006. 10th IEEE Singapore International Conference on, pages 1–5, Oct. 2006. [20] A. Warrier and I. Rhee. Stochastic analysis of wireless sensor network mac protocols. Technical report, 2005. [21] A. Woo and D. E. Culler. A transmission control scheme for media access in sensor networks. In MobiCom ’01: Proceedings of the 7th annual international conference on Mobile computing and networking, pages 221–235, New York, NY, USA, 2001. ACM. [22] E. Xiaopeng Ma; Lloyd. An incremental algorithm for broadcast scheduling in packet radio networks. Military Communications Conference, 1998. MILCOM 98. Proceedings., IEEE, 1:205–210 vol.1, 18-21 Oct 1998. [23] W. Ye, J. Heidemann, and D. Estrin. An energy-efficient mac protocol for wireless sensor networks, 2002.
[1] ANSI/IEEE. Wireless lan medium access control (mac) and physical layer(phy) specifications, 1999. [2] K. Arisha, M. Youssef, and M. Younis. Energy-aware tdmabased mac for sensor networks. IMPACCT 2002, May 2002. [3] M. Dawande, S. Gandham, and R. Prakash. Link scheduling in sensor networks: Distributed edge coloring revisited. IEEE INFOCOM 2005, March 2005. [4] M. Duflot, L. Fribourg, and C. Picaronny. Randomized dining philosophers without fairness assumption. In TCS ’02: Proceedings of the IFIP 17th World Computer Congress TC1 Stream / 2nd IFIP International Conference on Theoretical Computer Science, pages 169–180, 2002. [5] S. Ergen and P. Varaiya. Pedamacs: power efficient and delay aware medium access protocol for sensor networks. Mobile Computing, IEEE Transactions on, 5(7):920–930, July 2006. [6] D. Ganesan, B. Krishnamachari, A. Woo, D. Culler, D. Estrin, and S. Wicker. Complex behavior at scale: An experimental study of low-power wireless sensor networks, 2002. [7] V. A. Kottapalli, A. S. Kiremidjian, J. P. Lynch, E. Carryer, T. W. Kenny, K. H. Law, and Y. Lei. Two-tiered wireless sensor network architecture for structural health monitoring. volume 5057, pages 8–19, 2003. [8] I. Kurtis Kredo and P. Mohapatra. Medium access control in wireless sensor networks. Comput. Netw., 51(4):961–994, 2007. [9] K. Langendoen. Medium access control in wireless sensor networks, 2007. [10] C. Lin. On-demand qos routing in multihop mobile networks. INFOCOM 2001. Twentieth Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE, 3:1735–1744 vol.3, 2001. [11] R. Lin, Z. Wang, and Y. Sun. Energy efficient medium access control protocols for wireless sensor networks and its state-of-art. Industrial Electronics, 2004 IEEE International Symposium on, 1:669–674 vol. 1, May 2004. [12] G. Lu, B. Krishnamachari, and C. Raghavendra. An adaptive energy-efficient and low-latency mac for data gathering in wireless sensor networks. Parallel and Distributed Processing Symposium, 2004. Proceedings. 18th International, pages 224–, 26-30 April 2004.
35
