Professional Documents
Culture Documents
Abstract— We consider the problem of joint end-to-end band- cation of resources, such as transmission opportunities, radio
width allocation and radio resource management in WiMax channels, and transmit powers to links so as to meet perfor-
single-carrier wireless networks. The design problem is posed as mance objectives for the end-to-end transport rates. Recently,
a utility maximization problem subject to link rate constraints
which involve transmission scheduling and power allocation. a wide variety of optimization methods have been suggested
Inspired by a centralized algorithm for solving the associated for computing the achievable performance of systems that
optimization problem, we proceed systematically in our develop- coordinate multiple layers in the networking stack (see e.g. [3],
ment of distributed resource allocation mechanisms. Contrary to [4] and the references therein). These centralized schemes are
the centralized algorithm, the proposed solution is distributed useful for gaining insight into the performance benefits of
and of low computational complexity, generates schedules of
finite length and with fixed time-slot durations, and acts on coordinating the different layers of the protocol stack, but are
local neighborhood information only. Although the final scheme quite far from the distributed resource management schemes
is suboptimal, we isolate and quantify the performance losses needed in practice. Centralized solutions tend to incur large
incurred and demonstrate strong performance in examples. communication overhead costs, introduce a single point-of-
failure (the network control node) and scale poorly with the
I. I NTRODUCTION
number of network nodes. Moreover, many of the optimal
Although the research on multi-hop wireless networks has approaches are computationally demanding to execute.
a long history with some operational systems already on the This paper presents a distributed approach to joint end-to-
market, emerging standards such as 802.11s and 802.16 have end bandwidth sharing and transmission schedule construction
the potential of allowing wireless networking on a broader for WiMax-SC mesh networks. Our approach falls into the
scale. The IEEE 802.16 family of standards [1] is exploited by framework of network utility maximization (e.g. [5], [6],
a non-profit industry trade organization, WiMax, whose aim is [7]), since we formulate the optimal network operation as
to develop standards-based devices and services for delivering the solution of an optimization problem and apply math-
wireless broadband connectivity to business and home users. ematical decomposition techniques to find distributed solu-
WiMax devices should offer a mobile, flexible and rapidly tion mechanisms. Noting that the standard approach of dual
deployable alternative to the current cabled networks for a decomposition can not be used for solving our problem,
range of deployment scenarios, spanning from long-range, we devise a novel mathematical decomposition method that
low-density, line-of-sight conditions to short-range non-line- allow us to incrementally construct a schedule that provably
of-sight deployments in cluttered urban environments [2]. converges to the optimal. A unique feature of our scheme is
The basic WiMax architecture combines subscriber stations the distributed construction and maintenance of a transmission
(user devices), base stations, and wired backhaul services to schedule of multiple time-slots. Based on this scheme, we
deliver high data-rate to the end-users. WiMax devices support derive a heuristic for constructing finite-length schedules with
two network topologies, point-to-multipoint and mesh: in fixed time-slot length, and describe the basic functionalities
point-to-multipoint configuration, the base station is the only that needs to be implemented at sources, queues, and transmit-
transmitter operating in the downlink while a set of subscribers ters. This scheme improves significantly over other schemes
share the medium during the uplink communication; in mesh proposed in the literature. On the methodological side, we
mode, on the other hand, traffic can be routed also directly proceed systematically from the performance limitations given
between subscribers. To increase coverage at relatively low by a centralized optimization based on global information
cost, the standard specifies both single-carrier as well as multi- to a totally distributed, but suboptimal, scheme that relies
carrier physical layer technologies. WirelessMAN-SCa and on local information only. We isolate the performance losses
WirelessHUMAN are the parts of the standard that define the incurred in each simplification step and demonstrate strong
single carrier modulation scheme, in which a common channel improvements over existing approaches.
is shared by all users through centralized or decentralized
channel access mechanisms. This is the focus of this paper. II. N ETWORK M ODEL
A critical problem for traffic engineering in these systems, Consider a single carrier WiMax network consisting of radio
and in multi-hop wireless networks in general, is the allo- units (nodes) located at fixed positions. Each node is assumed
to have infinite buffering capacity and the ability to transmit, Thus, link l may transmit at rate rtgt if the SINR level at its
receive and relay data to other nodes. At any given time, nodes receiver exceed the threshold, and stays silent otherwise. Note
can either transmit or receive data from at most one other node. that a rate allocation with more states might be applied in order
We represent the network topology by a directed graph, with to get an higher resource utilization from finer granularity. This
nodes labelled n = 1, . . . , N and links labelled l = 1, . . . , L. leads to a finite number of achievable link-rate vectors
A link is represented by an ordered pair (i, j) of distinct nodes,
where the presence of link (i, j) means that the network is c(k) = (r1 . . . rL ) k = 1, . . . , K
able to send data from node i to node j. Nodes are assumed where rl ∈ {0, rtgt }. Although K may be as large as 2L , most
to always have data to send to the other nodes, possibly via networks will, due to interference and other technological con-
multi-hop routing. We label the source-destination pairs by straints, support substantially fewer link rate vectors. By time-
integers p = 1, . . . , P and let sp denote the end-to-end rate for sharing between a given set of link rate vectors {c(k) | k ∈ K},
communication between source-destination pair p. Associated we can achieve the following polyhedral rate region
with each pair p is a utility function up (·), which describes ( )
the utility of the pair to communicate at rate sp (cf. [8]). We K
X
(k)
X
C = c= αk c | αk ≥ 0, αk = 1
assume that up is increasing and strictly concave, with up →
k∈K k∈K
−∞ as sp → 0+ . Each demand is assumed to be routed along
a single fixed path between source and destination. The routing Here, the time-sharing coefficients αk represent the fraction
is specified by a routing matrix R = [rlp ] ∈ RL×P where of schedule length in which rate vector c(k) is activated. If
( C K contains all feasible rate-vectors, we will simply drop the
1 if data between pair p is routed across link l superscript and use the short-hand notation c ∈ C to denote
rlp =
0 otherwise that c is an achievable long-term average link rate.
Letting cl denote the transmission rate of link l, c = [cl ] III. N ETWORK - WIDE CROSS - LAYER OPTIMIZATION
be the vector of link-rates, the vector of total traffic across
A mathematical formulation of distributed end-to-end flow
links is given by Rs and the network flow model imposes the
control over TCP/IP networks has been developed in [8], [5].
following set of constraints on the end-to-end rate vector s
It is argued that the optimal network operation solves the
Rs c s0 network utility maximization problem
P
where the link rates depend on the medium access scheme, maximize p up (sp )
channel conditions and the allocation of radio resources, such subject to Rs c (3)
as transmit powers and time-slots, to the transmitters. s0
In single-carrier WiMax networks, all transmitters share the
where the variables are collected in the end-to-end rate vector
same frequency band, so interference will occur when multiple
s, while the link capacity vector c is assumed to be fixed.
links try to transmit at the same time. To model this, let Glm
If the utility functions are logarithmic, the problem yields a
be the effective link gain between the transmitter of link m and
proportionally fair allocation of end-to-end bandwidth [8].
the receiver of link l (including distance-based attenuation and
A distributed solution to this problem can be derived using
fading as well as the effects of coding gain, spreading gain and
mathematical decomposition techniques and suggests simple
beam-forming, see e.g., [9], [3]), let σl be the thermal noise
resource allocation mechanisms implemented in end-hosts and
power at the receiver of link l and Pl be the power used by
routers [8]. These functionalities can be mapped onto the
its own transmitter. We assume that each transmitter is subject
idealized operation of TCP clients and AQM algorithms, and
to a simple power limit 0 ≤ Pl ≤ Pmax , and define the signal
it has been argued that running the appropriate TCP and AQM
to noise and interference ratio of link l as
protocols effectively amounts to letting the network solve the
G P
γl (P) = P ll l (1) utility maximization problem [5]. In this case, congestion
σl + m6=l Glm Pm signals such as queue lengths can be interpreted as dual
where P = P1 · · P
· PL denotes the vector of power variables to the optimization problem (3). Our interest is on
allocation, and Il = distributed end-to-end flow control over wireless networks,
m6=l Glm Pm is the interference ex-
perienced at the receiver of link l. We view each link as a where the links capacities are not fixed a priori, but depend
single-user Gaussian channel with Shannon capacity cl (P) = on the allocation of radio resources. In the spirit of (3), the
W log(1 + γl (P)), where W is the system bandwidth. We associated utility maximization problem can be formulated as
assume that a link is able to transmit reliably when its SINR
P
maximize p up (sp )
is over a predefined threshold γ tgt , which also defines a unique subject to Rs c (4)
link rate rltgt = W log(1 + γ tgt ). To be able to see this effect c∈C s0
we introduce the following rate allocation policy
tgt where the optimization variables are the end-to-end rates s and
r , if γl (P) ≥ γ tgt the radio resource management parameters influencing c (i.e.,
rl = (2)
0, otherwise the power allocation and transmission scheduling).
A. Centralized optimization via column generation Step k
New slot under
negotitation
A centralized solution to (4) based on a column generation
scheme has been proposed in [3]. Starting from an initial set Step k+1
90
90
80
80
Total utility
Total utility
70
70
60
60
50
Optimal solution
Optimal solution TDMA scheme
50
Cross−Decomposition, incremental schedule 40 Cross−Decomposition
Cross−Decomposition, rolling horizon 2L Rolling horizon with distributed scheduling, 2L
Cross−Decomposition, rolling horizon 3L Rolling horizon and distributed power control
40 30
0 50 100 150 200 250 300 350 400 450 0 50 100 150 200 250 300 350 400 450
Iterations Iterations
Fig. 3. Utility as function of schedule length for the centralized approach, Fig. 4. Performance of combined rolling-horizon scheduling and greedy
cross decomposition, and rolling horizon scheduling. The scheduling subprob- transmission group formation.
lem is solved to optimality using global network information.
Scheme Total utility %
Optimal solution 95.88
A. Cross-decomposition and rolling horizon scheduling Rolling horizon 2L 94.80
−−
−1.13
The baseline for our evaluations is the performance of Cross decomposition 93.45 −2.53
Rolling horizon with distr. power control 88.33
the centralized optimization (4) based on global network Rolling horizon with distr. sched. 2L 81.33
−7.87
−16.22
information. While the optimal solution assumes variable time- TDMA scheme 57.07 −40.49
slot lengths, the techniques we propose in this paper construct
finite-length schedules with time-slots of fixed duration. Fig- TABLE I
ure 3 demonstrates the performance of a centralized approach, P ERFORMANCE LOSSES OF ROLLING HORIZON .
the cross-decomposition approach, and the rolling horizon
scheduling for schedule lengths of 2L and 3L respectively.
In these simulations, the scheduling subproblem is solved
to optimality using global knowledge accounting for the in- performance improvements of this scheme are significant.
terference when forming the transmission group (as in [3]). Similar evaluations have been made for a large set of networks
The rolling horizon scheduling show more rapid performance with comparable results.
improvements, since old transmission groups (generated with
VI. C ONCLUSIONS
outdated system information) are purged from the schedule.
However, due to the finite schedule length and fixed time-slot We have proposed a distributed solution to end-to-end
duration, the solution cannot achieve the performance of the bandwidth sharing and radio resource management in WiMax
centralized scheme. Although these effects are reduced as the single-carrier wireless networks. By posing the problem as
schedule length increases, a performance difference persists a utility maximization problem subject to link rate constraints
(compare the discussion around Figure 2). which involve both transmission scheduling and power alloca-
tion, we have proceeded systematically in our development of
B. The complete solution transparent distributed resource management schemes that fit
The complete solution combines the rolling horizon sched- into the WiMax standard. In the process, we have introduced a
ule construction with the greedy distributed transmission novel decomposition method for convex optimization and es-
group formation. The result for the same network scenario tablished its convergence for the utility maximization problem,
is shown in Figure 4. In this case, we can no longer see and combined this with a greedy transmission group formation
the rapid performance improvements from the rolling horizon scheme. Although the final solution is suboptimal, we have
scheduling under optimal transmission group formation, and isolated and quantified the performance losses incurred in each
the distributed scheme stabilizes with a distinct performance simplification step and demonstrated strong improvements
loss compared to the optimal solution. However, the perfor- over the state-of-the art solutions.
mance improvements over a TDMA scheme is substantial, and
some performance loss should be expected since the WiMax R EFERENCES
solution does not account for interference in the transmission [1] IEEE Standard for Local and Metropolitan Area Networks. Part 16: Air
group formation. For comparison purposes, we have evaluated Interface for Fixed Broadband Wireless Access Systems.
the performance of a distributed scheme that accounts for [2] A. Ghosh, D.R. Wolter, J.G. Andrews, and R. Chen. Broadband wireless
access with wimax/802.16: current performance benchmarks and future
interference in the transmission group formation (see [12] potential. In IEEE Communixations Magazine, 2005.
for details). In this case, the transmission group formation is [3] M. Johansson and L. Xiao. Scheduling, routing and power allocation
performed using a variant of the distributed power control with for fairness in wireless networks. In Proc. of IEEE VTC Spring, 2004.
[4] X. Lin and N. B. Shroff. The impact of imperfect scheduling on cross-
active link protection [15], which is not easily implementable layer rate control in wireless networks. In Proc. of the IEEE Infocom,
in the current standard. As can be seen in Figure 4, the Miami, FL, March 2005.
[5] S. Low and D. Lapsley. Optimization flow control I: Basic algorithm Lemma 1.4: The optimal Lagrange multipliers λ⋆ for the
and convergence. IEEE Trans. Networking, 7(6):861–874, 1999. capacity constraint in the definition of ν(c), with c ≻ ρ ≻ 0,
[6] L. Xiao, M. Johansson, and S. Boyd. Simultaneous routing and resource
allocation via dual decomposition. In Proc. of the 4th Asian Control are bounded and belong to a convex and compact set λ(ρ).
Conference, 2002. Proof: Follows analogously to Problem 5.3.1 in [16].
[7] M. Chiang. Balancing transport and physical layers in wireless multihop Lemma 1.5: The optimal Lagrange multipliers λ⋆ for the
networks: Jointly optimal congestion control and power control. IEEE
J. Selected Areas in Communications, 2005. capacity constraint defining ν(c) are unique.
[8] F. P. Kelly, A. K. Maulloo, and D. K. H. Tan. Rate control for com- Proof: The Karush-Kuhn-Tucker (KKT) conditions are
munication networks: shadow prices, proportional fairness and stability.
Journal of Operations Research Society, 49(3):237–252, March 1998. u′ (s⋆ (c)) = RT λ⋆ (c)
[9] S. Toumpis and A. Goldsmith. Capacity regions for wireless ad hoc Rs⋆ (c) = c
networks. IEEE Transaction Wireless Comm., 2(4):736–748, July 2002.
[10] J. T. Van Roy. Cross decomposition for mixed integer programming. The latter expression holds by the assumption that all links
Mathematical Programming, 25:46–63, 1983.
[11] F. Murphy. A column generation algorithm for nonlinear programming. are bottlenecks. The observation that the source rate s⋆ (c)
Mathematical Programming, 5(1):286 – 298, 1973. is unique since the objective functions are strictly concave
[12] Pablo Soldati. Distributed cross-layer optimized flow control for together with the fact that R has full row rank implies that
STDMA wireless networks. Master’s thesis, KTH, 2004.
[13] A. Muqattash and M. Krunz. CDMA-based MAC protocol for wireless λ⋆ (c) is unique.
ad hoc networks. In Mobihoc’03, 2003. Theorem 1.6: ν(c) is differentiable for all c ∈ C, c ≻ ρ ≻
[14] B. Hajek and G. Sasaki. Link scheduling in polynomial time. IEEE 0, and the derivative is λ⋆ (c).
Transactions on Information Theory, 34(5):910–917, September 1988.
[15] N. Bambos, S. C. Chen, and G.J. Pottie. Channel access algorithms with
Proof: ν(c) can now be expressed as
X
active link protection for wireless communication networks with power ν(c) = min max up (sp ) − λT (Rs − c) = min g(c, λ)
control. IEEE/ACM Trans. on Networking, 2000. λ∈λ(ρ) s≻0 λ∈λ(ρ)
[16] D. Bertsekas. Nonlinear Programming. Athena Scientific, 1999. p
[17] J. C. Dunn. Conditional gradient algorithms with open loop step-size where the set λ(ρ) is compact (Lemma 1.4), the optimal
rules. J. Mathematical Analysis and Applications, 62:432–444, 1978.
Lagrange multipliers λ⋆ (c) are unique (Lemma 1.5), and
A PPENDIX g(c, λ) is continuous and concave in c for all λ ∈ λ(ρ)
(Lemma 1.2). Furthermore, for every optimal Lagrange mul-
Convergence of the Cross-Decomposition Algorithm
tiplier λ⋆ , g(·, λ⋆ ) is differentiable at c, i.e.,
It is possible to prove convergence of the algorithm under
∂g(c, λ⋆ ) ∂ X
some assumptions, where the most restrictive assumptions are = up (s⋆p ) − λ⋆T (Rs⋆ − c) = λ⋆i
that all links are bottlenecks, the routing matrix has full row ∂ci ∂ci p
rank, and that the scheduling subproblem can be solved to Now Danskin’s theorem [16] gives the desired result.
optimality. The first assumption can be fulfilled by requiring Remark 1: Since a ρ fulfilling c ≻ ρ ≻ 0, can be found
that all links have at least one flow only using that link. for all c ∈ C, c ≻ 0, the theorem above gives that ν(c) is
First we consider some basic properties of ν(c) defined as differentiable for all c ∈ C, c ≻ 0.
X Lemma 1.7: The derivative of ν(c) is continuous
ν(c) = max up (sp )
Rsc, s≻0 Proof: Using the KKT conditions and the implicit func-
p
tion theorem, it can be shown that λ is differentiable with
Also define the dual function, g(c, λ), as respect to c. This means that λ is continuous in c, and the
X derivative is therefore continuous in c. See [16] for details.
g(c, λ) = max up (sp ) − λT (Rs − c) Theorem 1.8: Let u⋆ be the optimal value of the centralized
s≻0
p
cross-layer design problem (4). Algorithm 1 converges in the
Lemma 1.1: For every positive capacity vector, there exists sense that limk→∞ ν(c(k) ) → u⋆ .
a strict interior point s̄ to (4) and strong duality holds. Proof: The update rule for c(k) can be re-written as
Proof: A strict interior point s̄ satisfying s̄ ≻ 0 and k 1
Rs̄ ≺ c can be constructed by setting the elements in s̄ to c(k+1) = c(k) + c′ = (1 − ωk )c(k) + ωk c′
k+1 k+1
cl where ωk = 1/(k + 1) and c′ is found as the solution to the
s̄p = min − ǫ > 0, p = 1, ..., P
l L congestion-weighted throughput problem
where ǫ is a small positive constant. Since a strict interior maximize λT c′
point exists, Slater’s condition for constraint qualification is subject to c ∈ C
satisfied, hence strong duality holds [16].
Since λ is a gradient of ν(c) at c, this is a conditional gradient
Lemma 1.2: g(c, λ) is concave and continuous in c for all
method with an open loop step-size rule [17] that satisfies
λ0
ωk
Proof: Follows since g(c, λ) is linear in c. ωk+1 = , ω0 = 1
Lemma 1.3: ν(c) is concave, and subgradient to ν(c) at ωk + 1
c is given by the optimal Lagrange multipliers, λ⋆ , for the Since λ⋆ is continuous and the domain is compact, the
capacity constraint in the definition of ν(c). derivative of ν(c)Pis uniformly continuous. By Theorem 1
in [17], limn→∞ p up (sp (cn )) = p u⋆p .
P
Proof: See [16].