On Code Design in Joint MAC Scheduling and Wireless Network Coding

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

25th Biennial Symposium on Communications

On Code Design in Joint MAC Scheduling and Wireless Network Coding


Raheleh Niati Carleton University Ottawa, Ontario, Canada rniati@sce.carleton.ca Amir H. Banihashemi Carleton University Ottawa, Ontario, Canada ahashemi@sce.carleton.ca Thomas Kunz Carleton University Ottawa, Ontario, Canada tkunz@sce.carleton.ca

AbstractIn this paper, we investigate the problem of network coding and media scheduling in wireless multihop networks. Unique characteristics of the wireless media, such as omnidirectional transmissions and destructive interference, as well as the limitations of wireless nodes to either transmit or receive at any given time, imply new code design constraints for wireless networks. In this work, following the approach of Sagduyu et al., given a sufcient set of conict-free link sets, we investigate necessary and sufcient conditions to design a capacity-achieving network coding solution for wireless networks. This approach uses a wired graph representation of the wireless network to design wireless network codes. Our study shows that the proposed approach works only if we preserve all of the code characteristics of the wired graph in the wireless one. We show the shortcoming of previous studies through a counterexample and explore this case further. The resulting insight is summarized in Theorem 1, which proves that unequal time shares allocated to conict-free link sets will result in changing coding coefcients as the sets with larger time share are scheduled repeatedly.

I. I NTRODUCTION Since Ahlswede et al. [1] introduced the notion of network coding in data communications networks, this research area received great attention from both communications and computer network research groups. Many fundamental results, based on wired networks, have been published in the recent years [2], [3]. However, networks that would benet substantially from network coding are multihop wireless networks. As wireless bandwidth is orders of magnitude lower than wired bandwidth, throughput is severely limited. Interference of ows across multiple hops along a shared wireless channel reduces the throughput even further. Unfortunately, not all of the results derived for wired networks are directly applicable to wireless networks. There has been some efforts to develop practical wireless network coding protocols [4], [5], but in our research we are more interested in fundamental results. In particular, we would like to explore how to apply results from the wired domain to the wireless domain, rather than deriving new theoretical results for wireless multihop networks. This approach has been shared by other researchers, with various degrees of success. Some early work fell short, as it only considered some of the unique aspects of wireless networks. For example, [6] uses articial nodes to model omnidirectional transmissions in wireless networks. This modication changes the wireless network to a link-based graph that can not handle MAC scheduling. In [7] it is shown that, from a practical point of view, network coding does not necessarily result in an improvement of throughput. The paper gives two

examples that show that independently designing network coding (i.e. the way information streams combine together in a given network) and MAC scheduling may in fact result in less throughput. On the other hand, joint scheduling and network coding results in higher throughput. In [8] it is shown that the general problem of interference-free scheduling for multihop wireless networks is NP-hard. Consequently, even though a joint design of network coding and scheduling improves network throughput, there is no guarantee that we achieve the optimal solution in polynomial time. The reason is that the problem of interference-free scheduling of a wireless network is equivalent to nding maximal independent sets of a graph that represents the wireless network. Edges that belong to the same independent set can be scheduled together because transmission from any of these nodes does not interfere with other nodes of the same independent set. Finally, [9], [10] are close to our goal and provide a solid starting point for our research. We keep the node-based representation of wireless networks introduced in this work, but identify a shortcoming as demonstrated through an example in the paper and extend the work to correctly deal with the identied problem. The general approach to joint scheduling and code design, as proposed in [9], [10], is a two-step process: 1. Build a group of interference-free network realizations through heuristics. A network realization is a partition of a subset of all network nodes into transmitters and receivers. The heuristic starts with assigning a random node (excluding the source) of the wireless graph as a receiver and nds a transmitter (if possible) with minimum cost (power) to reach that node. Receivers and transmitters are added to a realization until we can add no more node to the realization without violating the interference-free condition. The same process will start over for the next realization by choosing another random node as receiver. The stopping criterion for this process is when each node (except the source node) is assigned to be a transmitter and receiver at least once in the total set of realizations. The heuristic implicitly chooses each wireless link at most once. 2. Allocate time fractions to wireless network realizations. After building a sufcient set of realizations according to the above heuristic, a time fraction is assigned to each of the network realizations. It species the duration of time (in one schedule period) that each realization can be active. These time fractions depend on the performance criterion of the network. Different possible performance criteria in a wireless

978-1-4244-5711-3/10/$26.00 2010 IEEE

456

25th Biennial Symposium on Communications

network are maximum ow, minimum delay, minimum cost (energy or power), and selecting any of these performance criteria may result in different time fractions for network realizations. The maximum ow criterion is one of the most straightforward criteria to calculate time fractions because, using the maximum ow-minimum cut theorem, we can solve for the unknown values of time fractions to achieve max ow using the procedure described in [10]. After determining the time fractions, the wireless graph is represented by a wired one with the same set of nodes and links of the network realizations. Edges in the wired graph have weights (capacities) that correspond to the wireless realization time fractions for those realizations in which an edge is active. In this step we have to make some assumptions to maintain/model the omnidirectionality of transmissions in wireless networks. These assumptions are briey discussed in Section III. Based on this wired presentation and additional conditions, network codes are designed using [11], [12]. Then, the wired network codes are used to determine the coefcients of the wireless network codes, as will be discussed in more detail in Section III. The key contribution of this paper is in an improved analysis and understanding of this last step. Using an example in Section IV, we show how the previous approach in [9], [10] does not always preserve code characteristics of wired networks in wireless networks. Our example studies the effect of having non-equal time fractions for network realizations. II. W IRELESS N ETWORK M ODEL AND P ROBLEM F ORMULATION After a brief review of the unique wireless network characteristics in Section II.A, a formal model is formulated in Section II.B and serves as the basis for the remainder of the paper. A. Wireless Network Characteristics & Model A wireless network has the following characteristics [13]: Activity of each wireless link depends on the current status of other transmission links. Interference from other links plays an important role in transmissions. Therefore MAC scheduling has a signicant inuence on congestion. The most important characteristic of wireless networks is the broadcast medium. Although a broadcast medium is generally known to be an advantage of wireless networks [14], [15], it is the main cause of the two previous properties, resulting in a signicant difference between wireless networks and wired ones. In addition, in the limiting case, wireless nodes only have a single radio. With current technology, these radios operate in half-duplex mode, i.e., they either transmit or receive. Therefore, our model of single-radio wireless nodes performs omnidirectional transmissions, has a single transceiver per node, and avoids destructive interferences [9], [10]. The general system model for a multi-hop wireless network is a directed acyclic graph (DAG) G = (V, E), where V is the set of all nodes and E denotes all edges. Undirected edges

can be replaced with two parallel edges of opposite directions. Also let S V be the set of independent sources and D V be the set of destinations. A wireless multicast problem is distributing some messages from sources si S to all the destination nodes dj D. For the wireless channel model, we consider a classical collision channel model with sharp boundaries for the radio propagation region around each wireless node. Transmissions from a node are deliverable only inside its transmission region, and other nodes can receive transmissions inside their own regions. Our model does not deal with physical locations of the nodes. We start with the connectivity graph of wireless nodes, which directly models links between any two nodes in the classical collision channel model. In the next section we formalize the wireless network coding problem. The notations and formulation are similar to those of [16], [9], and [10]. B. Problem Formulation Wireless characteristics of a network require time-varying network ows. Because wireless nodes should be scheduled in different timeslots to avoid interference, coding coefcients of a wireless node are not necessarily xed during the time. Transmissions in a wireless graph have to be scheduled for independent sets of the graph. A time fraction is assigned to each network realization (one independent set) to transmit its messages. Therefore, the time varying ows and their respective constraints are dened as follows. We consider a total of f M realizations for a network: N f = Nm , m = 1, 2, ..., M . f f Each realization Nm has its own set of nodes Vm and links f f Em and it is active during time fraction m . Vm is composed of two disjoint sets of T (m) and R(m) , which denote sets f of transmitters and receivers in network realization Nm such (m) (m) f that T R = Vm . Therefore we have the following constraints for any d D, i V, j V and m = 1, 2, ..., M
xi,j (d) and zi,j {0, 1} , xi,j (d) zi,j
(m) (m) (m) (m)

(1)

(m) f For the Classical Channel Model: (i, j) Em , zi,j = 1 if

Pi

(m)

Pi,j , Pk

(m)

< Pk,j k = i & Pj

(m)

=0

(2)

Flow conservation constraint:


M m=1

m
j:(i,j)Em
f

(m) xi,j (d)

j:(j,i)Em
f

(m) xj,i (d)

=r

(3)

if i = s, r if i = d or 0 otherwise (m) In these relations, zi,j is the total ow on link (i, j) during (m) f time fraction m when realization Nm is active. xi,j (d) is the portion of that ow on the same link destined to node d D. (m) Pi is the power cost for transmitting messages from node i and Pi,j is the power required by node i to reach node j in a single hop of the classical collision channel model. Starting from constraint (2), the required set of realizations is built. Then based on constraints (1) & (3), any problem with its own performance criteria can be solved. The solution is the value (m) of xi,j (d)s and time fraction m s. However, condition (3) is nonlinear. As explained in Section I, and in the approach used by [10], if there exists a relationship between the performance criteria, ows and m s, then we can start by determining time

457

25th Biennial Symposium on Communications

fraction values and solve a Linear Program to determine the ows. The LP would be: maximize r subject to constraints (1) & (3) We can take advantage of the fact that the minimum value of all cuts is equal to the maximum possible ow in the network. Therefore, predetermining m values results in a Linear Program for ow values. To get the maximum ow of a wireless graph, we rst determine cut values as a function of m s and x their values. Then, because the maximum ow of a network is equal to its minimum cut, the ows are easily determined. Given these ows and time fractions values, capacity-achieving network codes can then be designed, as discussed in the next section. III. R EVIEW OF W IRELESS N ETWORK C ODE C ONSTRUCTION In [11], Koetter & Medard proposed an algebraic approach to construct linear network codes that achieve the capacity of the network. Their formulation is used for wired networks because the network coefcients they introduced are all linkbased. It is assumed that there are a total of (v) discrete and independent random processes at each node v of the network. These random processes are going to be delivered and replicated in the network to be received by a group of nodes. In this way, for any link of the network, e=(v,u) such that u = head(e) & v = tail(e) there exists a random process, called Y(e) that satises
(v)

messages at a wireless node:


Y (n) (v) =
m

(
u:(u,v)Em (v) (m,n)
f

(m,n) ((u, v))Y (m) (u)

(6)

+
j=1

(v)Xj

(m)

(v))

Y (n) (v) is the encoded ow at node v and during realization (m,n) f Nn . It uses the coding coefcients (m,n) ((u, v)) and j that weigh the incoming streams from node v and the j th source stream at node u respectively. Also, the k th output f ow is decoded at node v V in network realization Nn as
Zk (v) =
m u:(u,v)E f m (n) (m,n) ((u, v))Y (m) (u) k

(7)

Y(e) =
l=1

e,l X(v, l) +
e:head()=tail(e) e

e,e Y () e

(4)

Coding coefcients e,l and e,e are elements of a nite eld F2m and X(v, l) is one of the total (v) source streams received at tail(e). The following relation shows the decoded process at any receiver node v
Z(v,j) =
e:head()=v e e,j Y

() e

(5)

e,j s are decoding coefcients and members of F2m and max(j) is the total number of independent random processes that are going to be decoded at node v. Figure 1-a shows the relationship between Xs and Y s. If we want to employ the same strategy to design wireless network codes, omnidirectional transmission of nodes has to be incorporated into the coding coefcients. In other words, network coding in wireless networks is called to be a node-based design rather than link-based. Relation (6) shows the linear combination of

Figure 1-b shows the relation between Xs and Y s. By comparing Figure 1-a & 1-b, the node-based versus link-based design of network codes for wireless and wired networks is clear. According to Lemma 1 in [10], the wireless network realizations (N f ) and the corresponding hypothetical wired network graph (N g ) have the same cut values and therefore have the same maximum ows. Therefore a capacity-achieving coding solution for N f achieves the capacity in N g as well, of course, after the solution is transfered from N f to N g . The claim is correct only if all of the coding characteristics of the wired network are preserved in the wireless network. The transformation introduced by [10] preserves the omnidirectionality of wireless nodes and assigns the wired coding coefcients to the realizations that occur in different time fractions. It is instructive to look at the rules of transforming a network coding solution from the wired network (N g ) to the wireless one (N f ). Relations (8-10) describe the mappings (m,n) (m,n) from j (.), (., .), and k (.) to j (.), j (.), and (m,n) (.). A more detailed description is presented in [10]. k Although these transformations are intended to maintain the wireless properties in the resulting wireless coding coefcients, they fail to deliver all properties of the link-based coding coefcients. We will describe this with an example in Section IV. The importance and necessity of properly transferring the wired coding solution to the wireless coding solution arises because otherwise there can be no guarantee that the resulting wireless coding solution will achieve the same (maximal) throughput as the coding solution for the wired network.
j
(m,n) j () e (m,n) f (tail(e)) = 1, if j (e) = 1, e En f Em , e

(8) (9) (10)

= 1, if (e, e) = 1, e = 1, if
k (e)

f En

and m = n

(m,n) (e) k

f = 1, e Em , and f = 1, e Ep

n = max p :

k (e)

Fig. 1.

Coding coefcients in a) wired networks, b) wireless networks

In the next part we present an example that shows a capacityachieving linear network code for N g does not necessarily achieve the capacity of N f if the transformation is performed using only the rules in [10]. In a transformation from wireless to wired graph with m capacities, if all of the m s are equal, then the links of the wired graph are exactly the same as those of the wireless graph. In this particular case, by choosing a sufciently large unit of time, all of the link capacities will be equal to 1 and we can build network codes using relations (6) & (7). However, in case of non-equal time fractions, some modications are required to maintain all properties of the link-based transmission of N g in the wireless graph.

458

25th Biennial Symposium on Communications

IV. U NEQUAL T IME F RACTIONS We start with a wireless multicast example shown in Figure 2. This network has 3 independent sources and 3 destinations. According to a proof by Ford & Fulkerson [17], if we have a network in which every source is willing to transmit to all of the destinations, then this network is equivalent to a multicast network that has a single source and a group of destinations. This is equivalent to having a virtual source node connected to sources {s1 , s2 , s3 } in 2. Figure 3 shows a sufcient set of network realizations of the example graph. There are a total of 5 sufcient realizations in this network. Solving an optimization problem (as discussed in Part II.B) to maximize the total rate gives the following results:
(1) = (2) = (3) = 1/7, (4) = (5) = 2/7 r = 3/7 = 0.42 packets/timeslot

Achieving the max possible rate (here r=0.42 packets/timeslot) requires using realizations 4 and 5 twice as often as the other three realizations. A schedule period of the wireless network in which every node transmits some combinations of its incoming streams includes one time share f f f allocated to each of realizations N1 , N2 , and N3 and two f f time shares to each of realizations N4 and N5 . Furthermore,

represented by G = (V, E), each edge e is assumed to have an integer edge capacity. By choosing a sufciently large unit of time, any network can be approximated arbitrarily accurately by a network with edges having integer capacities [12]. After presenting all edge capacities by integers, capacities larger than 1 are modelled by parallel edges between two nodes. Now if we build a wired graph based on the capacities (time fractions) that we determined above, Figure 4-a shows the resulting graph. A capacity-achieving network coding solution for the wired graph is shown in Figure 4-b. For this specic wired graph, there is no capacity-achieving coding scheme such that the coding coefcients for the two outgoing edges of u are the same. To avoid ambiguity we do not talk about all of the network coding coefcients in both networks. Wireless coefcients are derived using relations (8-10). Now if we focus on two of the edges in N g dened by: e7 = e8 = (u, v) we see that based on relation (4), random processes leaving node u in the wired network are:
Y (e7 ) = (e4 , e7 )Y (e4 ) + (e5 , e7 )Y (e5 ) Y (e8 ) = (e5 , e8 )Y (e5 ) + (e6 , e8 )Y (e6 )

According to relation (6), the stream out of node u in the wireless graph will be the following
Y (4) (u) = (1,4) (e7 )Y (1) (s1 ) + (2,4) (e7 )Y (2) (s2 ) + (2,4) (e8 )Y (2) (s2 ) + (3,4) (e8 )Y (3) (s3 )

Because wireless coding is a node-based scheme rather than link-based, it does not differentiate between e7 and e8 . When the coding coefcients of the wired graph are transferred back to the wireless one, it represents a broadcast model which does not preserve the properties of the wired coding scheme. Knowing that (2,4) (e7 ) = (2,4) (e8 ) = 1 and the coefcients are in F2 , the stream out of node u will be:
Y (4) (u) = (1,4) (e7 )Y (1) (s1 ) + (3,4) (e8 )Y (3) (s3 ) Fig. 2. Wireless Network Example

Fig. 3.

Network Realizations

Accordingly, nodes d1 and d3 can receive information from the rst and the third sources during 5 time fractions and node d2 is only able to receive information from the second source in the same time duration, because it is unable to decode the rest of the messages. Consequently, the receiving rate at nodes d1 and d3 is 2 = 0.4 packets/timeslot and the receiving rate 5 at node d2 is 1 = 0.2 packets/timeslot. None of these rates 5 are the maximum achievable rate of the wireless graph. In g order for N and N f to have the same rate, nodes in N f should perform the same encoding and decoding operations as determined for the corresponding wired network. The approach presented in [10] works for wireless graphs in which all of the time fractions are equal. Otherwise, we have to preserve the following condition while wired coding coefcients are transferred to wireless ones: If there are two (or more) edges like ei and ej in the wired graph such that tail(ei )=tail(ej )=u, and they are part of the same network realization, but their coding coefcients are different, they belong to repetitions of a network realization. Therefore, messages transmitted from node u during the repetitions of a realization should be specied separately. In our example graph, we can specify messages on edges e7 and e8 of the wired graph in the following two wireless linear combinations transmitted from node u during the two repetitions of the fourth realization:
Y1
(4)

the prerequisites of network code design says in a network

(u) = (1,4) (e7 )Y (1) (s1 ) + (2,4) (e7 )Y (2) (s2 )

459

25th Biennial Symposium on Communications

Y2

(4)

(u) = (2,4) (e8 )Y (2) (s2 ) + (3,4) (e8 )Y (3) (s3 )

Table I shows a network coding solution with scheduling (4) (4) based on these Y1 (u) and Y2 (u) that allows the network to transmit different linear combinations when a realization is repeated. In this table s1 b1 u means s1 transmits b1 to u. The total ow of this solution is 0.42 packets/timeslot and is equal to the optimal rate.

based on heuristics because the general problem of multicast media scheduling is NP-hard. The overall goal of our work is how to transform a wireless graph into an equivalent wired graph, determine an optimal coding solution for the wired graph, and transfer the result back into a coding solution for the wireless graph that achieves the same performance. We showed through an example that a previously developed approach does not always result in optimal solutions. We developed a new condition that guarantees the wireless network will achieve the same rate as the wired network. It deals with the conditions that conict-free network realizations have different time shares as a result of a linear optimization problem. We proved that when we have unequal time shares for network realizations, coding coefcients of network realizations with larger time shares will change as these realizations are scheduled repeatedly within a cycle. For future work, we plan to have a closer look at joint MAC scheduling and network coding in wireless networks and how to design necessary and sufcient realizations for wireless graphs. R EFERENCES
[1] R. Ahlswede, C. Ning, S.Y.R. Li and R. W. Yeung, Network Information Flow, IEEE Transactions on Information Theory, Vol. 46, No. 4, pp. 1204-1216, 2000. [2] S. Y. R. Li, R. W. Yeung, and N. Cai. Linear network coding, in IEEE Transactions on Information Theory, Feb. 2003. [3] R. Koetter, M. Medard, An Algebraic Approach to Network Coding, in IEEE/ACM Transactions on Networking, Vol. 11, No. 5, Oct. 2003. [4] S. Katti, H. Rahul, W. Hu, D. Katabi, M. Medard, and J. Crowcroft, XORs in The Air: Practical Wireless Network Coding, In ACM SIGCOMM 2006, Pisa, Italy, Sep. 11-15, 2006. [5] S. Katti, D. Katabi, Wenjun Hu, and Rahul Hariharan, The Importance of Being Opportunistic: Practical Network Coding For Wireless Environments, In Proc. 43rd Allerton Conference on Communication, Control, and Computing, Monticello, IL, Sep. 2005. [6] Y. Wu, P. A. Chou, Q. Zhang, K. Jain, W. Zhu, and S. Y. Kung, Network Planning in Wireless Ad-Hoc Networks: A Cross-Layer Approach, IEEE J. Sel. Areas Commun., Vol. 23, No. 1, pp. 136-150, Jan. 2005. [7] P. Chaporkar, and A. Proutiere, Adaptavie Network Coding and Scheduling for Maximizing Throughput in Wireless Network, MobiCom07, Sep 9-14, Montreal, Quebec, Canada, pp. 135-146. [8] K. Jain, J. Padhye, V. N. Padmanabhan, and L. Qiu, Impact of Interference on Multi-hop Wireless Network Performance, MobiCom03, Sep 14-19, 2003, San Diego, California, pp. 66-80. [9] Y. E. Sagduyu, and A. Ephremides, Joint Scheduling and Wireless Network Coding, First Workshop on Network Coding, Theory and Applications, Riva del Garda, Italy, Apr. 2005. [10] Y. E. Sagduyu, and A. Ephremides, On Joint MAC and Network Coding in Wireless Ad Hoc Networks, IEEE Transactions on Information Theory, Vol. 53, No. 10, Oct. 2007. [11] R. Koetter, and M. Medard, Beyond Routing: An Algebraic Approach to Network Coding, in IEEE/ACM Transactions on Networking Volume 11, Issue 5, Oct. 2003, pp. 782-796. [12] S. Jaggi, P. A. Chou, and K. Jain, Low Complexity Algebraic Multicast Network Codes, in Proc. ISIT, June 2003. [13] Y. Wu, P. A. Chou, Q. Zhang, K. Jain, W. Zhu, and S. Y Kung, Network Planning in Wireless Ad Hoc Networks: A Cross-Layer Approach, IEEE Journal on Selected Areas in Communications, Vol. 23, No. 1, Jan. 2005. [14] C. Fragouli, and E. Solganin, Network Coding Applications, now Publishers Inc., 2007. [15] C. C. Wang, and N. B. Shroff, On Wireless Network Scheduling with Intersession Network Coding, in Proceedings of the 42th Conference of Information Sciences and Systems, Princeton, USA, March 19-21, 2008. [16] D. S. Lun, M. Mdard, T. Ho, R. Koetter, Network Coding with a Cost Criterion, in Proc. 2004 International Symposium on Information Theory and its Applications (ISITA 2004), Oct. 2004, pp. 12321237. [17] L. R. Ford, D. R. Fulkerson, Flows in Networks, Princeton University Press, 1962. pp. 15-17.

Fig. 4. a) Wired counterpart graph, b) Capacity achieving coding coefcients for the wired network

Timeslot Schedule Timeslot Schedule

1 s1 b1 u 5 v(b1 b2 )D

2 s2 b 2 u 6 u(b2 b3 )v

3 s3 b 3 u 7 u(b2 b3 )D

4 ub1 b2 v -

TABLE I N EWTORK CODING SOLUTION FOR THE WIRELESS NETWORK IN F IG . 3

So far, we examined the effect of having non-equal time fractions when we solve a linear programming problem to optimize a criterion of interest in the wireless networks through an example. Theorem 1 generalizes the (potential) impact of such non-equal time fractions on network codes to achieve the equivalent performance in both networks. Theorem 1: The equivalent wired graph of a wireless graph will have parallel edges with different coding combinations, if and only if solving an LP optimization problem (of interest) for the set of conict-free realizations of the wireless graph results in non-equal time fractions. Proof: (forward direction) Suppose i is the time fraction of nodes in realization i and i = m , where m is a constant and is the smallest time fraction of all of the realizations in the network. Now assume for every node in realization i, there are less than m outgoing messages with different coding coefcients. Suppose that the maximum number of different coding coefcients out of all nodes in realization i is m (m < m). Therefore, all of the nodes in realization i can transfer their messages during m duration of time. It means m is not part of the optimal solution. (backward direction) Suppose there are m different encoding combinations to be delivered by one wireless node (lets say u) in realization i. Also assume that there exists at least one node v that is interested in all of the m different combinations from node u. These coded streams must belong to realization i, because all of them are transmitted over link (u, v). Therefore, realization i must happen m times during one schedule period. V. C ONCLUSIONS In this work, we studied the problem of wireless network coding and media scheduling. Similar to others, our work is

460

You might also like