of the major energy consumers, energy efficient forwarding algorithms are essential to provide an infrastructure for any kind of application. Taking advantage of ...
Energy-Efficient Aggregation Forwarding for Wireless Sensor Networks Marcel Busse, Thomas Haenselmann, and Wolfgang Effelsberg Computer Science IV - University of Mannheim, Seminargeb¨aude A5 D-68159 Mannheim, Germany {busse, haenselmann, effelsberg}@informatik.uni-mannheim.de
Abstract Based on an energy efficiency metric that is defined by the fraction of end-to-end delivery ratio and required energy, we propose a new distributed algorithm for the case of correlated data. In contrast to related work, our metric considers energy savings due to aggregation a priori and achieves a superior performance. However, a problem that arises from the ability to aggregate and to change energy cost during the construction of a forwarding tree is that routing cycles may occur because information used may be out-dated. By employing different types of sequence numbers, our strategy is able to prevent such cycles, accounting for the distributed nature of sensor networks.
1. Introduction As the radio transceiver of wireless sensor nodes is one of the major energy consumers, energy efficient forwarding algorithms are essential to provide an infrastructure for any kind of application. Taking advantage of data aggregation further reduces the communication overhead, especially if it is taken into account during the establishment of forwarding paths. By in-network processing, data gathering in a sensor network is commonly processed based on aggregation trees, which can be considered as reverse multicast trees rooted at a network sink. Inner nodes of the tree first summarize received data and then forward the aggregated data to their parent nodes until the sink is reached. Therefore, special aggregation function like SUM, COUNT, AVG and MIN/MAX, or even more advanced functions like approximations for medians can be used [12]. The remainder of this paper is structured as follows: The next section briefly outlines related work. Models and assumptions are discussed in Section 3. Section 4 presents an energy-efficient forwarding strategy that is extended by exploiting in-network processing in Sections 6 and 5. Results obtained by means of simulations as well as a real-world evaluation are discussed in Section 7 and Section 8. Section 9 concludes the paper.
2. Related work A first proposal for data-centric routing is directed diffusion [5], which is extended in [4] in order to facilitate data aggregation based on greedy increment trees. TTDD [9] improves the directed diffusion approach by exploiting a two-tier grid structure based on geographic information, which is used to restrict flooding to particular regions. Similarly, HEED [14] divides the network into clusters, with assigned cluster heads to perform aggregation. The problem of double-counting data, which may occur in multi-path routing, is solved by Nath et al. in [10], using the concept of synopsis diffusion. Boukerche et al. [1] propose an aggregation tree construction where leaf nodes go into a lowpower sleep mode to save energy, while other nodes stay awake to participate in data dissemination. Similar to [3], the number of leaves is maximized during the tree construction in order to save the most amount of energy. In [7], Lee and Wong present E-Span, an algorithm for energy-aware spanning trees that are rooted by the source node with the highest residual energy. Parent nodes are chosen by selecting the best energy node along the shortest path towards the source. After gathering all data, the root of the tree then forwards an aggregated packet to the sink. In [8], E-Span is extended by a lifetime-preserving tree (LPT). Here, the tree is rooted by the source node that has the highest socalled tree energy, which is defined by the minimum residual energy of all non-leaf nodes. Each node selects its parent such that the minimum residual energy on the path towards the root is maximized. Jia et al. [6] have proposed GIST, a group-independent spanning tree that considers the case of unknown sources. Their aim is to find a single optimum spanning tree that can be used for all subsets of source nodes. The idea is to built the tree only once and use it for any group of data sources. The tree is constructed by using a hierarchical approach, dividing the network into several grids. A location-based quad-tree is then established by assigning each grid cell a leader. The leader of the entire network is the sink node, which is also the root of the quadtree. As data is always sent to the leader of the upper level, aggregated packets finally arrive at the network sink.
3. Models and assumptions We model the packet reception ratio (PRR) on a wireless link according to an extended log-normal shadowing model. For d being the distance between two arbitrary nodes, the log-normal behavior is defined by 1 d 2α d 1|εpoll 1 − ρi,n ∀j∈Ωi e + prri,j (Ej + εack ) , (4)
with ρi,k = ∀j∈Ωi ,α(j)≤k (1 − prri,j prrj,i ), and , ack , and poll being the required energy to transmit data packets, acknowledgements and polling messages, respectively. The MEEF strategy then locally maximizes the energy efficiency Eief f defined as Eir /Eie for node i by examining each neighbor j. In case of n = 1, we get the SEEF strategy since only a single neighbor is considered. Please refer to [2] for more details and a deeper understanding.
5. Energy-efficient aggregation forwarding As constructing an aggregation tree is similar to the Steiner tree problem which is NP-hard, we simplify the problem to the case where only source nodes are able to aggregate data packets (although aggregation may be later performed by any node in the network). This simplification allows us to construct a minimum spanning tree approximation, with path cost between two sources calculated according to the energy efficiency metric used in the previous section. While the end-to-end delivery ratio Eir for a node i remains the same, the required energy Eie changes if the information issued by node i are aggregated along the forwarding path. Hence, in case i is a source node, its energy cost Eie is set to zero. Like in SEEF and MEEF, Eir and Eie are then periodically propagated to adjacent neighbors in order to construct the aggregation tree. The extended forwarding strategies are called single-link (SEEAF) and multi-link energy-efficient aggregation forwarding (MEEAF).
5.1. Routing cycles
5.2. Algorithm description
In order to avoid routing cycles, we use the following heuristic: Unlike in SEEF and MEEF, each node only considers neighbors that have a better end-to-end delivery ratio than the node itself. That is, rather than using the energy efficiency, the end-to-end delivery ratio serves as a monotonically decreasing function. In addition, we use a hop counter to break ties. Thus, at time t a node i calculates its energy efficiency Eief f by only considering neighbors j that satisfy
The forwarding algorithm is shown in Algorithm 1. Assuming an arbitrary node i receives a forwarding update message from a neighbor j, which contains information regarding Ejr , Eje , Ejh , and sj , i selects a single forwarder (SEEAF) or a set of forwarding nodes (MEEAF) by calculating its energy efficiency. If j is a direct parent, node i first updates its sequence numbers si and s maxi (line 2). It then checks if j is a new neighbor or if the forwarding variables of j have changed since the last update. If so, the information concerning the forwarding path of j is updated (line 5). Additionally, i recalculates its energy efficiency (lines 6-18) by considering all potential forwarding sets. However, due to the conditions in line 8 and 10, only nonchild nodes are taken into account. After each forwarder set has been examined, the one that maximizes the node’s energy efficiency is stored for later use, together with Eir , Eie , and Eih . At last, i updates its sequence number s changei (line 20) in order to exclude out-of-date neighbors the next time, in case its parent(s), the end-to-end delivery ratio or the energy cost of its forwarding path have changed.
r r h (Ei,t−1 < Ejr ) ∨ ((Ei,t−1 = Ejr ) ∧ (Ei,t−1 > Ejh )) (5)
and r r h (Ei,t < Ejr ) ∨ ((Ei,t = Ejr ) ∧ (Ei,t > Ejh )),
(6)
with Eih being the expected forwarding path length, which can be calculated similar to Equation 4 by
h j∈Ωi ρi,α(j)−1 prri,j (Ej + 1) h
. (7) Ei = j∈Ωi ρi,α(j)−1 prri,j However, routing cycles may still occur if out-dated information is used. Thus, sequence numbers are employed to ensure that the information received from neighbors is up-to-date. The general idea is that the sink node periodically generates an increasing sequence number, which is propagated throughout the network, however, only from a parent node to all its children. Each time the forwarding variables of a node i change, the node sets a timestamp that is equal to the current node’s sequence number. Hence, the next time the node calculates its energy efficiency, it only considers neighbors that satisfy requirements 5 and 6 and in addition have a higher sequence number than the one set at the last node’s change. Therefore, three different sequence numbers are needed: si , s changei , and s maxi . The sequence number si is propagated by a node to all its children. Each time a node receives such a sequence number from a parent node j it updates its own sequence number by setting si = sj . However, since in the multi-link case each node might have several parents, the node needs to set its sequence number to si = min {sj } ∀parents j
(8)
instead. The maximum sequence number is set to s maxi = max {s maxi , sj }, ∀parents j
(9)
each time node i receives a sequence number sj from one of its parents. Using the maximum sequence number node i has ever seen, s changei is then set to s maxi if i’s forwarding variables change or a reset occurs. With these different sequence numbers on hand, we can formulate the third requirement a neighbor j of i must satisfy by requiring sj ≥ s changei .
(10)
6. Greedy increment tree Another way to tackle the Steiner tree problem is to use a greedy approach in a centralized fashion. At first, no node in the network is aware of aggregation, and a simple shortest path algorithm is carried out, with the link cost being the energy efficiency according to Equations 3 and 4. The aggregation tree is then constructed by adding the best source node concerning its efficiency to the still empty tree (together with all nodes along the appropriate path). In the next step, each node calculates its shortest path to the sink again. However, this time, nodes belonging to the aggregation tree cause no extra costs. Thus, the energy efficiency is computed by using Equation 4 with zero energy cost if i already belongs to the aggregation tree and Eie otherwise. After each node has determined its new energy efficiency, the next source node not yet covered by the tree is added. This procedure continues until all source nodes belong to the aggregation tree. Due to the greedy property of the algorithm, the tree is called greedy increment tree (GIT).
7. Simulations We simulated the following forwarding strategies: SEEF, MEEF, SEEAF, MEEAF, GIT, LPT [8], and GIST [6]. While LPT was implemented as described in [8], GIST was slightly modified in order to achieve a better performance. Therefore, the quad-tree construction was changed to take energy efficiency into account, i. e., as in EEF and EEAF, packets are forwarded to a destination along the best
Algorithm 1 Receive EEAF message(j, Ejr , Eje , Ejh , sj ) 1: if j ∈ parentsi then 2: si ← sj , s maxi ← max{s maxi , sj } 3: end if 4: if change in j’s forwarding variables then 5: store j’s new forwarding variables ˆ r∗ ← 0, E ˆ e∗ ← ∞, E ˆ h∗ ← 0 6: E i i i 7: for all forwarding sets Ωi do 8: if ∀ˆ j ∈ Ωi : (sˆj ≥ s changei ) ∧ ((Eir < Eˆr ) ∨ ((Eir = j
9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: 22:
Eˆr ) ∧ (Eih > Eˆh ))) then j j ˆr , E ˆe, E ˆ h according to Eq. 3, 4, 7 calculate E i i i r ˆ ˆ r = E r ) ∧ (E ˆ h > E h )) ˆ if ∀j ∈ Ωi : (Ei < Eˆr ) ∨ ((E i i ˆ ˆ j j j then ˆ r /E ˆe > E ˆ r∗ /E ˆ e∗ then if E i i i i ˆ r∗ ← E ˆr , E ˆ e∗ ← E ˆe, E ˆ h∗ ← E ˆh E i i i i i i set forwarders according to Ωi end if end if end if end for ˆ e∗ , E r ← E ˆ r∗ , E h ← E ˆ h∗ Eie ← (i is a source node) ? 0 : E i i i i i if forwarders or forwarding variables changed then s changei ← s maxi end if end if
energy-efficient paths, however, using intermediate nodes according to the established quad-tree.
7.1. Node density The influence of the node density concerning different performance metrics is shown in Figure 1 for a source fraction α of 20%. Each data point illustrates the average over 500 simulation runs. Except for the LPT strategy, all strategies achieve a high ratio of delivered information as shown Figure 1(a), which expresses the accuracy of delivered aggregation values. The delivery ratio of LPT is around 0.25 and thus not shown. The GIST strategy slightly suffers from sending data packets along its quad-tree, which causes longer forwarding paths. In contrast, MEEF and SEEF establish direct forwarding paths. Thus, their path lengths are shorter, which, in this case, leads to better delivery ratios. However, because EEAF and GIT take advantage of data aggregation they are able to establish more short-distance links that show better reception characteristics, without worsening the energy efficiency of forwarding paths. Hence, the more source nodes exist, the higher the ratios of delivered information. Since multi-link forwarding improves the delivery performance of forwarding nodes in the majority of cases, both MEEAF and MEEF outperform SEEAF and SEEF, respectively. As additionally the number of data retransmissions decreases if multi-link forwarding is applied, the average energy cost of each node decreases, too. Figure 1(b) shows the average energy consumption in the network of all strate-
gies. GIT performs best due to its greedy construction characteristics. However, the energy cost of EEAF is only slightly higher, considering the fact that it works in a fully distributed way. While the performance of LPT is worst because much energy is consumed for retransmitting data packets, EEF and GIST perform similarly. For higher node densities, GIST slightly benefits from the established quadtree as aggregation takes place, early. However, since the quad-tree is group-independent, it is clearly outperformed by MEEAF as well as SEEAF. The trade-off between the information delivery ratio and the energy consumption as expressed by the energy efficiency is depicted in Figure 1(c). The best efficiency is achieved by GIT. However, GIT is not expected to be applicable in large sensor networks consisting of hundreds of nodes. As Figure 1(c) shows, among all distributed algorithms, SEEAF as well as MEEAF perform best and almost reach the energy efficiency of GIT forwarding. Although SEEAF is outperformed by MEEAF due to better delivery ratios and additionally lower energy cost, it shows considerable improvements over EEF and GIST. Interestingly, GIST does not achieve a significantly better energy efficiency in our simulation settings than MEEF or SEEF, although forwarding along the quad-tree saves energy. However, the energy savings comes at the expense of lower delivery ratios. As LPT suffers from bad delivery ratios as well as a high energy consumption, it leads to the worst energy efficiency.
7.2. Fraction of source nodes In addition to different node densities, we have also varied the source fraction from 10% to 100% for a fixed density µ of 30 nodes. The results are depicted in Figure 2. Figure 2(a) shows the delivery ratio of information issued by source nodes. The delivery ratio of LPT is again about 0.25 and thus not shown. As the number of sources does not influence the forwarding tree construction of EEF, GIST, and LPT, their delivery ratios is almost constant, even if the source fraction increases. In contrast, GIT and EEAF are able to improve their performance because more source nodes lead to more aggregation points within the network, which is beneficial for both algorithms. Thus, if the number of source nodes increases, sending packets towards a nearby aggregation path, which has a better end-to-end delivery, becomes more efficient. Since EEF as well as LPT and GIST neglect such aggregation gains during the construction of forwarding paths, longer but better forwarding paths concerning their delivery performance are avoided due to unconsidered energy savings. Although the total energy consumption of the network grows for a higher source fraction due to more issued packets, the average energy consumption on forwarding paths decreases as shown in Figure 2(b). Thus, the higher the number of sources, the greater the aggregation gain per
0.7 0.6
0.9 0.85 0.8 MEEAF SEEAF MEEF SEEF GIT LPT GIST
0.75 0.7
0.5
1.2 MEEAF SEEAF MEEF SEEF GIT LPT GIST
Energy Efficiency [bits/µJ]
0.95 Consumed Energy [mJ]
Information Delivery Ratio
1
0.4 0.3 0.2 0.1
0.65
0 10
15
20
25 30 35 Node Density
40
45
50
1 0.8 0.6 MEEAF SEEAF MEEF SEEF GIT LPT GIST
0.4 0.2 0
10
(a) Information delivery ratio
15
20
25 30 35 Node Density
40
45
50
10
15
(b) Consumed energy
20
25 30 35 Node Density
40
45
50
(c) Energy efficiency
0.7 0.6
0.98 0.97 0.96 MEEAF SEEAF MEEF SEEF GIT LPT GIST
0.95 0.94
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.5 0.4 0.3 0.2 0.1
0.93 0.1
2 MEEAF SEEAF MEEF SEEF GIT LPT GIST
Energy Efficiency [bits/µJ]
1 0.99 Consumed Energy [mJ]
Information Delivery Ratio
Figure 1. Influence of the node density (α = 0.2, R = 3)
0.9
Fraction of Source Nodes
(a) Information delivery ratio
1
0 0.1
1.8 1.6 1.4 1.2 MEEAF SEEAF MEEF SEEF GIT LPT GIST
1 0.8 0.6 0.4 0.2 0
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.1
0.2
0.3
Fraction of Source Nodes
(b) Consumed energy
0.4
0.5
0.6
0.7
0.8
0.9
1
Fraction of Source Nodes
(c) Energy efficiency
Figure 2. Influence of the source fraction (µ = 30, R = 3) source node. As already illustrated in Figure 1(b), LPT consumes the most amount of energy due to a very high number of required retransmissions. SEEF, MEEF, and GIST perform similar, achieving almost the same energy consumption as EEAF and GIT for α → 1. However, as Figure 2(c) shows, EEF, GIST, and LPT are outperformed by SEEAF as well as MEEAF considerably, independent of the number of source nodes within the network. As in Section 7.1, the GIT strategy gives an upper bound concerning the energy efficiency, which is almost reached by MEEAF for α → 1. In conclusion, the simulation results indicate that the forwarding metrics used by SEEAF and MEEAF are quite reasonable, showing the desired network performance if the source nodes are known a priori. Otherwise, EEF is a good alternative as it still performs better than GIST and LPT.
starting randomly. The maximum number of retransmissions is set to three. The evaluation lasts for 200 rounds (100 minutes) such that each source sends 200 data packets in total. After sources are randomly selected, the network sink starts sending periodic beacons, triggering the aggregation tree construction. To simplify matters, link measurements are carried out before in order to estimate packet reception ratios. Upon establishing all forwarding paths, the actual evaluation starts. Data packets are issued by source nodes and sent to the first forwarding node, which does not forwarded them immediately but buffers the data until the next round. In doing so, several data packets are aggregated and forwarded to the next node at once.
8. Experimental evaluations
Results are presented in Table 1, which shows average values taken over 10 runs. The first row presents the energy efficiency of each strategy, confirming our simulation results from Section 7. As before, by establishing a greedy increment aggregation tree (GIT), the best performance is reached. Although about 6% worse, MEEAF achieves the best efficiency among the remaining strategies. SEEAF slightly performs worse than MEEAF, but yet better than SEEF, MEEF, LPT, and GIST. Compared to EEF, EEAF improves the energy efficiency of about 6% to 7%. Similar
We also carried out real-world experiments, evaluating the performance of aggregation in our WSN testbed consisting of 25 ESB [11] sensor nodes placed in a grid structure with one fixed sink node. All nodes use a transmission power of 15% and establish forwarding paths according to the strategy being evaluated. Among all nodes, five nodes are picked randomly, which act as sources issuing 32 byte data packets at a rate of one packet per round (30 seconds),
8.1. Evaluation results
Table 1. Experimental evaluation results (EEAF) Strategy Energy Efficiency [bits/mJ] TX Data TX Control RX Data RX Control Delivered Information Units [%] Total Energy Consumption [mJ] Hop Counter Retransmissions
MEEAF 0.842 98.33 84.19 86.29 110.78 99.31 320.08 2.87 0.21
SEEAF 0.833 101.44 75.03 83.86 87.56 99.04 323.07 2.96 0.22
to the results we obtained by simulations, the LPT strategy performs worst, consuming more than twice the energy spent by EEAF and EEF due to significantly more packet transmissions (TX data). The fact that LPT uses worse forwarding links is shown by the number of retransmissions. While packets in LPT travels almost half the number of hops compared to the other strategies, the number of retransmissions per link is more than five times higher. LPT thus achieves a delivery ratio of about 55% only; while all other strategies achieve a delivery ratio of almost 100%. The total number of transmitted data packets averaged over all nodes is shown in the second row of Table 1. Again, multi-link forwarding reduces the number of transmitted data packets compared to single-link forwarding. At the same time, more control packets need to be sent due to the polling of backup nodes (TX control). Thus, although less data packets are sent, more packets are received by forwarding nodes (RX data and control). Furthermore, as indicated by the hop counter, backup nodes are often closer to the sink. Polling those nodes increases the likelihood of forwarding data successfully, reducing the energy consumption due to less retransmissions. In total, the network energy consumption caused by a source node is between 60 and 75 mJ; for LPT about 180 mJ. Thus, except for LPT, all strategies achieve an energy efficiency of more than 0.78 bits/mJ.
9. Conclusions In this paper, we analyzed the impact of data aggregation based on energy-efficient forwarding (EEF). By taking energy savings due to aggregation into account a priori, we extended EEF and further improved the overall energy efficiency. EEAF is easy to implement, but care must be taken in order to avoid routing cycles. We thus proposed a simple set of sequence numbers which is used to identify out-dated information. The simulations as well as the experimental evaluations have shown that EEAF outperforms EEF clearly concerning delivery ratios, consumed energy, and energy efficiency. For future work, we will extend EEAF by considering partially correlated data, too. Also, long-time experiments using real applications are planned.
MEEF 0.796 112.86 100.90 105.06 130.88 98.94 356.64 2.66 0.29
SEEF 0.78 113.93 90.19 100.12 101.08 98.96 363.63 2.70 0.31
GIT 0.894 95.92 74.40 84.40 84.47 98.98 305.99 3.30 0.21
LPT 0.183 297.14 65.14 164.14 85.46 55.33 896.43 1.71 1.63
GIST 0.781 109.56 76.68 85.64 92.56 99.14 348.35 2.63 0.26
References [1] A. Boukerche, X. Cheng, and J. Linus. Energy-aware datacentric routing in microsensor networks. In Proc. of MSWIM, San Diego, CA, Sep. 2003 [2] M. Busse, T. Haenselmann, and W. Effelsberg. Energyefficient forwarding schemes for wireless sensor networks. In Proc. of WOWMOM, Niagara-Falls, NY, Jun. 2006. [3] M. Ding, X. Cheng, and G. Xue. Aggregation tree construction in sensor networks. In Proc. of VTC, Orlando, FL, Oct. 2003. [4] C. Intanagonwiwat, D. Estrin, R. Govindan, and J. Heidemann. Impact of network density on data aggregation in wireless sensor networks. In Proc. of ICDCS, Washington, DC, Jul. 2002. [5] C. Intanagonwiwat, R. Govindan, D. Estrin, J. Heidemann, and F. Silva. Directed diffusion for wireless sensor networking. Trans. on Networking, 11(1):2–16, Feb. 2003. [6] L. Jia, G. Noubir, R. Rajaraman, and R. Sundaram. Gist: Group-independent spanning tree for data aggregation in dense sensor networks. In Proc. of DCOSS, San Francisco, CA, Jun. 2006. [7] M. Lee and V. W. S. Wong. An energy-efficient spanning tree algorithm for data aggregation in wireless sensor networks. In Proc. of PACRIM, Victoria, BC, Aug. 2005. [8] M. Lee and V. W. S. Wong. E-span and lpt for data aggregation in wireless sensor networks. Computer Communications, 29(13–14):2506–2520, Mar. 2006. [9] H. Luo, F. Ye, J. Cheng, S. Lu, and L. Zhang. A two-tier data dissemination model for large-scale wireless sensor networks. Wireless Networks, 11(1–2):161–175, Jan. 2005. [10] S. Nath, P. B. Gibbons, S. Seshan, and Z. R. Anderson. Synopsis diffusion for robust aggregation in sensor networks. In Proc. of SENSYS, Baltimore, MD, Nov. 2004. [11] J. Schiller, A. Liers, H. Ritter, R. Winter, and T. Voigt. ScatterWeb - Low power sensor nodes and energy aware routing. In Proc. of HICSS, Maui, HI, Jan. 2005. [12] N. Shrivastava, C. Buragohain, and D. Agrawal. Medians and beyond: New aggregation techniques for sensor networks. In Proc. of SENSYS, Baltimore, MD, Nov. 2004. [13] A. Woo, T. Tong, and D. Culler. Taming the underlying challenges for reliable multihop routing in sensor networks. In Proc. of SENSYS, Los Angeles, CA, Nov. 2003. [14] O. Younis and S. Fahmy. Heed: A hybrid, energy-efficient, distributed clustering approach for ad-hoc sensor networks. Trans. on Mobile Computing, 3(4):366–379, Oct. 2004. [15] M. Zuniga and B. Krishnamachari. Analyzing the transitional region in low power wireless links. In Proc. of SECON, Santa Clara, CA, Oct. 2004.
