Optimal Routing, Link Scheduling and Power Control in Multi-hop Wireless Networks R. L. Cruz and Arvind V. Santhanam University of California, San Diego IEEE INFOCOM 2003

Outline Motivation Objective Optimal Scheduling and Power Control Routing and Capacity Allocation Resource Allocation Examples Conclusion

Motivation Minimizing the total power in broadband wireless networks is of paramount importance : Increase operational lifetime in the case of battery powered devices Coexist symbiotically with other systems which share the same frequency spectrum. Most currently deployed networks that support high data rates use only a “single hop” for wireless communication However, use of multiple hops to transport data has been shown to enhance network capacity and may be necessary due to cabling limitations in many environments

Network Set Up Network of base-stations interconnected to each other through wireless links. Each base-station serves as an ingress or egress for the aggregate of traffic associated with the mobile users in its domain Each base-station routes its data through other base stations via multiple wireless hops to an access point connected to a wired infrastructure.

Objective Joint routing, link scheduling and power control to support high data rates for broadband wireless multi-hop networks Finding an optimal link scheduling and power control policy that minimizes the total average transmission power in the wireless multi-hop network, subject to given constraints regarding the minimum average data rate per link, as well as peak transmission power constraints per node. Incorporating a Globally optimum routing scheme.

Notations N stationary nodes (base stations) 1,2,…N A set ε of Lε transmission links, among the possible N(N −1) links between nodes, constitutes a network topology. Link l(i,j) : transmitter node i uses signal power P(l) G(i,j) – path gain from node i to node j (Assumed to be constant) T(l) : Transmitting node of link l R(l) : Receiving node of link l ηj : Ambient noise power at node j : SNR for link l X(l) : Achieved data rate on link l W : Frequency bandwidth

Shannon Capacity of link = Assumption : X(l) is a linear function of (At small values of SIR) All links share the same frequency band of width W. Assuming the Gaussian approximation to compute the BER, the data rate of a link l with a tolerable BER of 10−q and using BPSK modulation is given by

C(l) : Desired data rate on link l. Thus Matrix form : If F is a stable matrix, i.e. all eigenvalues of F are strictly inside the unit circle, then (I − F)−1 exists and has nonnegative elements. Thus,

Foschini-Miljanic algorithm for cellular network Constraints: can be seen as the Minimal power vector that supports the network topology Foschini-Miljanic algorithm for cellular network Constraints: If eigenvalues of F are on or outside the unit circle, there does not exist any power vector P which supports the required data rates. If the eigenvalues of F are inside the unit circle but close to the boundary, the minimal power vector will be very large. Even if the eigenvalues of F are well inside the unit circle, using the minimal power vector (I – F)-1 b may be inefficient.

Example A square network of 4 nodes and 2 links {(1, 2), (3, 4)}. G(1, 2) = G(3, 4) = G0 and G(1, 4) = G(3, 2) = G0 /2 X({1, 2}) = X({3, 4}) = W’/2 n0 = G0 (at both receivers) (I – F)-1 b = (0.67 0.67)T But, if time sharing is employed between the 2 transmitters, and each transmitter active for 50% of the time -> the same average data rate of W’/2 is achieved, with average power of 0.5 Watt. It can be inefficient to operate with all links in a wireless network active concurrently However, in order to achieve high data rates with constraints on peak transmission power, concurrent transmissions may be necessitated

C(l) are determined by the average rates at which traffic is generated by users and the routing algorithm that is used. In wireless networks it is possible to reconfigure the data rates of the links on a fast time scale in response to changing traffic and channel conditions.

Optimal Scheduling and Power Control (1) Time divided into slots. Transmissions begin and end on slot boundaries Pm(l) : Transmission power for the transmitter T(l) for link l in slot m be the network power vector for slot m P max (i) : maximum transmission power of node i : links in ε that originate at node i

Optimal Scheduling and Power Control (2) Constraint on Peak Transmission Power : (1) Achieved data rate for link l in slot m Long term average rate of link l Constraint on Minimum acceptable average data rate : (2)

Optimal Scheduling and Power Control (3) Required minimum average rate vector Average power consumed by the transmitter for link l Average network power vector There may or may not exist a sequence of network power vectors P1, P2, . . . that satisfy (1) and (2). If there does exist a sequence of such network power vectors, aim is to minimize a linear function of where α(l) is a positive weight

Primal Problem subject to constraints (1) and (2) (3) This optimization involves choosing optimal power levels in each slot for each transmitter

Duality Approach (1) Let the optimal cost in (3) as a function of be denoted as if there is no network power vector satisfying (1) & (2) is a convex function of the vector Set of Dual variables Potential function V as Dual objective function is defined as subject to (1) For any non-negative vector subject to (1) & (2)

Duality Approach (2) Dual Optimization Problem (4) The components of the optimal dual variable vector represent a sensitivity of the optimal cost with respect to a perturbation in the minimum average data rate for a link. Thus, (Optimization over a single slot) where

Duality Approach (3) Let M be the number of extreme points of SP denote the extreme points of SP . Each point in SP can be represented as a combination of Upper bound on M is 2Lε It can be shown that Complexity of computing is O(M) is an affine function of . Thus, Iterative ascent algorithm used to solve the dual problem (4)

Feasibility check If is feasible, then If is infeasible.

Computing the Optimal Policy Solving (4) yields the optimal dual variable vector and extremal power vectors P*,i such that (5) If is finite, an optimal schedule of network power vector vectors exists that consists solely P*,i Thus, in the optimal policy, every node is transmitting at possible peak power to exactly one receiver, or not transmitting at all. where K be the number of extremal network power vectors P*,i such that (5) holds. ,

Optimal Policy Assuming is finite, Xavg(l) is exactly equal to C(l) in an optimal policy Let X*,i (l) : Rate of link l corresponding to the P*,i Assuming is finite, there exists a “weight vector” such that indicates the relative frequency at which the extremal network power vector P*,i is utilized in an optimal policy

Reducing Complexity Optimal schedule can have no more than (Lε + 1) transmission modes M can be as large as 2Lε Complexity of minimizing dual objective function can be exponential Reducing the set of possible transmission modes considered can greatly reduce complexity. No node transmits or receives data at the same time in the optimal schedule For multihop networks, transmission modes which consist of multiple simultaneous transmissions to a receiver are eliminated

Hierarchical Link Scheduling and Power Control Hierarchical approach to minimize the total average transmission power of all the links in a network with a large number of links Links in the network partitioned into groups called clusters Links in a cluster geographically close to each other Links in one cluster scheduled somewhat independently of links in other clusters Cluster level scheduling is done at the top Clusters geographically far way from each other impose negligible mutual interference If desired data rate on links are sufficiently low, the optimal policy activates all the clusters in the network simultaneously.

Routing and Capacity Allocation Optimal values for the dual variables . Can estimate the cost of supporting additional traffic on each link If an additional ε units of traffic to be routed along route r, then where r(l) = 1 if link l is included in route r and r(l) = 0 otherwise Shortest path algorithm with weights for link l to the sensibility Since H(.) is convex, increases after additional traffic is allocated on link l. Thus, routes initially unattractive may become more attractive after traffic is added on other links. Optimal paths does not always correspond to minimum energy paths.

Resource Allocation Examples

String Topology 1 1 – Source of data 5 - Sink Each node has omni-directional antenna with peak power constraint of 1 Watt. Ambient nose power at all nodes assumed to be constant

String Topology (2) W’ = 107 Optimal policy schedules concurrent transmissions even though they are in close geographic proximity

String Topology (3) Below 4 Mbits/sec, scheduling policy reducesto TDMA. Between 4 Mbits/sec and 4.98 Mbits/sec, [{(1, 2)},{(2, 3)}, {(3, 4)} and {(1, 2), (4, 5)}]. above 4.98 Mbits/sec : Optimal transmission modes are[{(1, 2)}, {(2, 3)}, {(3, 4)}, {(1, 2), (3, 4)} and {(1, 2), (4, 5)}].

Diamond Topology Node 1 only source of data, and node 4 is the sink Peak transmission power each node is fixed at 1 Watt G(1,4) = 1/ d(1,4)4 . Other path losses are given by inverse square law of distance. Minimum energy path 1 -> 2 -> 4 W’ = 107

Diamond Topology (2) Traffic split over multiple paths even for moderate levels of ambient noise

Diamond Topology (3) Traffic loads below 3.65 Mbits/sec, optimal policy essentially TDMA Beyond 3.65 Mbits/sec, optimal policy is by scheduling {(1, 2), (3, 4)} and {(1, 3), (2, 4)} for a dominant fraction of time and scheduling transmission modes {(1, 2)}, {(2, 4)} for the remaining time.

Hierarchical Topology Nodes 1,2 and 3 are the sources of data and node 6 is the sink G(1, 6) = 1 / d(1,6)4 and G(2, 6) = 1 / d(2,6)4 The minimum energy paths for the source nodes are: {1 3 6}, {2 3 6} and {3 6}

Hierarchical Topology (2) For low ambient noise, optimal transmission modes {(1, 3)}, {(2, 3)} and {(3, 6)} according to TDMA but also allows modes{(1, 4), (5, 6)} and {(2, 5), (4, 6)} to be active for reasonable fractions of time. For sufficiently high ambient noise, dominant transmission modes are {(1, 4), (5, 6), (2, 3)} and {(2, 5), (4, 6), (1, 3)} and {(1, 4), (2, 5), (3, 6)}.

Minimizing Total Average Transmitter and Receiver Power New cost function to minimize : where θ(l) is the energy expended per bit that is constant for each receiver node R(l). Accounting for receiver power in formulation changes the choice of routes and the rate allocations on links in each route. Using multiple hops in such topologies would consume higher receiver energy than using a single hop, making multihop routing inefficient.

Conclusion The integrated routing, scheduling and power control framework supports higher throughputs at the expense of decreased energy efficiency. Framework well suited for slow fading wireless channels which are relatively constant for long durations of time. Time synchronization needed between transmitters Framework applicable for providing a means for wireless interconnection of fixed stationary nodes, like wireless access nodes In optimal policy, each nodes is either transmitting at peak power to a single receiver or not transmitting at all, which it simple to implement. Energy could be conserved by “sleep” schedules The optimal link schedule time-shares a small number of optimal subsets of links (Lε + 1) in order to achieve the required data rates.

Conclusion (2) Optimal policy for low required data rates or in low ambient noise regimes schedules links in a TDMA sequence. In moderate noise regimes, or to achieve higher data rates, optimal strategy involves scheduling multiple simultaneous transmissions even though they may be in close geographic proximity. Minimal required average power is a convex function of the required data rates on each link. As the level of ambient noise increases, non-minimum energy paths can be exploited to increase throughput, despite the fact that all links share a common bandwidth.