Beruflich Dokumente
Kultur Dokumente
Abstract—Advances in device technology and dense wavelength or weeks. This difference in time scale arises primarily from
division multiplexing have enabled modern optical backbones to the common lack of both support and applications: lack of
provide high data rate services in a cost-effective manner. At support for dynamic traffic in commercial networks hinders ap-
the same time, applications are emerging with the potential to
leverage these capabilities for short time spans: load migration plication development, while lack of widespread applications
for Internet clouds, integrated computational grids, and remote limits incentive for support in commercial networks. Interest
surgery. Until such applications mature, investment in real in rapid provisioning dates back at least a decade [7], and use
networks to support them is unattractive, but without such of a dynamic network may save up to 80% capacity [8] relative
investment, the applications mature slowly. Current networks to a supporting dynamic traffic with a static network. The need
thus remain dominated by lightpaths provisioned by hand, over
the space of days or weeks. for on-demand wavelength connection services for future IP
Three problems must be solved to enable a rapid transition customers has also been growing [9], [10]. The emergence
to dynamic optical networks that integrate well with demands at of relatively short duration applications that require high data
the Internet Protocol (IP) level: traffic characterization, resource rates for video delivery, health care applications, etc. [11],
dimensioning, and resource allocation. Traffic characterization has increased the dynamic variability of traffic. Migration of
provides models of traffic demand variation and evolution, which
can only be predicted at this stage. load between cloud computing sites, dynamic peer-to-peer and
In this paper, we provide an overview of our past and current distributed computing applications, and computational science
work on the resource dimensioning and allocation subproblems. applications executing across the globe (grids) can all generate
We describe our methodology for dimensioning networks based high data-rate demands over short time periods. Once support
on expected traffic behavior. We next offer a few observations
for dynamic wavelengths becomes stable and cost efficient,
about dealing with statistical variations in network behavior. We
develop analytic models of opportunity cost and a congestion- demand will grow rapidly.
aware routing algorithm based on the models. A near-optimal The challenges for realizing the vision of a dynamic optical
threshold value for this algorithm can be chosen based on core are threefold: traffic characterization, resource dimen-
analysis, and simulations demonstrate that this choice is robust sioning, and resource allocation. Traffic characterization is
to variations in both topology and capacity.
the process of developing models that can be used to predict
variations and evolution of traffic demands. Resource dimen-
I. I NTRODUCTION sioning is the problem of provisioning hardware resources in
Global optical backbones and core networks rely primarily a way that minimizes cost while providing a low likelihood
on Dense Wavelength Division Multiplexing (DWDM) [1]. of overload (e.g., using a blocked call metric). Resource
Advances in technology for fast switching, frequency tuning, allocation is the problem of mapping dynamic connection
and control plane integration have enabled these networks requests to hardware resources.
to provide high data-rate services in a cost-efficient man- In the absence of existing dynamic demands, traffic char-
ner [2], [3]. For example, Reconfigurable Optical Add-Drop acterization is a prediction problem. A Poisson traffic model
Multiplexers (ROADM) and MicroElectroMechanical (MEM) is used for connection request arrival and departures. Aggre-
switches have gained popularity in core networks [4], [5]. gation at lower levels generally suppresses bursty behavior
Generalized Multi-protocol Label Switching (GMPLS) [6] in backbone networks [12]; although dynamic wavelength
provides signaling capability to directly reserve wavelengths services may alter this behavior, little or no data is available.
on optical devices for protocols such as IP, SONET, and ATM. In this paper, we describe our efforts over the last few
Actual use of these networks, however, remains dominated years in developing mechanisms for the resource dimensioning
by lightpaths provisioned by hand and over the space of days and allocation problems. We focus on an Optical-Electronic-
Optical network (or equivalently, an optically transparent
The material presented in this paper is based in part upon work supported by
National Science Foundation grant ANI 01-21662 ITR. The authors are also network with full wavelength conversion at each node) with
grateful for resources provided through generous donations by Intel as well centralized control for routing. Wavelength capacity on each
as support from the Information Trust Institute of the University of Illinois at
link is the primary resource. Lightpaths are automatically set
Urbana-Champaign and the Hewlett-Packard Company through its Adaptive
up (torn down) by OXCs upon the acceptance (departure) of
Enterprise Grid Program. The content of this paper does not necessarily reflect
the position nor the policies of any of these organizations. connection requests. The connections are bidirectional and
978-1-4244-7596-4/10/$26.00 ©2010 IEEE
2
TABLE I
use the same path in both directions. Based on the rout- N ETWORK NOTATION .
ing policy and the current availability of wavelengths, the N Set of nodes.
network either accepts or rejects a connection request. No E Set of links.
queuing/preemption/rerouting is assumed. Ce Total capacity of link e ∈ E.
Ue Current used capacity of link e ∈ E.
Backbone topologies are constrained by geography, and R Set of all end-to-end request pairs. R ⊆ N × N .
abundant fibers are laid to avoid future digging. Costs thus κr Mean requested capacity for pair r.
consist primarily of those associated with operating wave- λr Mean arrival rate for pair r.
μr Mean departure rate for pair r. The expected holding
lengths, such as the number of line cards and amplifiers. time is 1/μr .
In this context, resource dimensioning focuses on allocating T SLr Topological shortest path length for pair r.
link capacities for a given topology. Our initial work [13], T SPr Topological shortest paths set for pair r.
T {λr , κr , μr } Traffic matrix with per pair arrival rate λr , capacity
[14] outlined a strategy for dimensioning networks based on κr and departure rate μr .
estimated traffic demands and targeted load, and evaluations η Congestion threshold.
of the impact of traffic demand evolution. Later work (not
Algorithm 1: Dimensioning for matrix T {λr , κr , μr }.
discussed in this paper) extends dimensioning to problems of
link failure restoration [15] and multiple domains [16]. 1 ∀e ∈ E, Ce ← 0 foreach request pair r ∈ R do
2 foreach topological shortest path p ∈ T SPr do
Resource allocation focuses on online routing. We assume 3 foreach link e ∈ p do
demands for full wavelengths, and thus ignore the problem 4 Ce ← Ce + μ λ|Tr κ r
SPr |
r
of traffic grooming. We cover highlights of route selection 5 end
strategy (details in [17]), then and develop analytic models 6 end
7 end
of opportunity cost that motivate a simple congestion-aware 8 Round each link capacity Ce to the nearest integer.
routing scheme with an admission control threshold parameter
that is robust across both topologies and network capacity [18],
[19]. Comparisons with oracular optimal routing and more
We dimension networks using Algorithm 1, a basic dimen-
detailed analyses of other algorithms appear in the papers.
sioning technique that allocates wavelength resources accord-
ing to topology and Poisson traffic demands. Our notation is
II. R ESOURCE D IMENSIONING summarized in Table I. The algorithm dimensions capacities
Previous studies on circuit switching networks tried to solve for each request pair uniformly over all topological shortest
the resource dimensioning and allocation problems jointly, paths for the pair. The initial wavelength capacity on each link
planning an optimal path (or a few alternative paths) for is dimensioned at 100% of the projected load. A fraction of
each expected connection demand. Models were proposed to the projected load is then used to model the actual offered
compute the blocking rate for connection requests [20], [21], load on a network. If the offered load increases, the network
becomes
congested. The projected load of a network is defined
[22], [23]. These approaches worked well for legacy low data-
λr κr
rate networks with limited resources but became insufficient as r∈R μr T SLr / e∈E Ce .
for today’s high data-rate models [2], [24], [11]. Separating
the two problems can lead to better performance [13]. III. R ESOURCE A LLOCATION
Adding fibers to an optical backbone network is clearly Interest in dynamic wavelength services has led to renewed
not feasible at time scales of minutes or hours. Even when interest in online, link-state based routing and load balancing
extra fibers are available, however, activation of these fibers between pre-optimized and dynamically selected paths [9],
typically requires installation of hardware such as transponders [10]. In this section, we illustrate the relationship between
and switches as well as time for tuning and testing the path length, network load, and blocking rate.
system. These operations currently take weeks or months. We evaluate dynamic routing algorithms in terms of the
Dynamic optical networks are thus unlikely to remain at an expected number of blocked requests for a given number
optimal offline design point between upgrades, and resource of arrivals. For the simulations in this paper, we allow the
dimensioning must address not only statistical variations for network to reach steady state (which generally requires about
current load, but evolution of that load over time to avoid the 20,000 arrivals), then measure blocking probability over 5,000
need for more frequent upgrades. arrivals. Each data point represents an average over many
The use of properly dimensioned networks is important traffic demands selected at random from a space of possible
when studying the behavior of dynamic resource allocation demands in which each node pair has a demand uniformly
schemes, as misdimensioned networks tend to obscure dif- distributed between 1 and 10 connection requests per unit time.
ferences. The key insight in this regard is that most misdi- Each matrix is normalized to the average load in the space of
mensioned networks will have some weak point, typically a possible demand matrices, and the network is dimensioned
small cut with too few resources to meet the demand across by averaging the results of Algorithm 1 across the randomly
the cut. As a result, only routing behavior across the cut will selected matrices. Samples are generated using a new set of
be evaluated; raising the load past the point at which the cut random matrices until the 95% confidence interval for each
saturates makes blocking unacceptably high (more than a few data point lies within 5% of its value. We report results only
percent). Networks with uniform capacity, for example, are for the ARPANET network; results for other networks are
most likely misdimensioned, and should be avoided if possible. similar [13], [14], [17], [18], [19].
3
z
0.1
0.01
threshold
Plane z=1 y
hc-TSL+any
hc-TSL+0 i/c
hc-TSL+1 2
0.001 nk
hc-TSL+2 Li
hc-TSL+3
0.75 0.8 0.85 0.9 0.95 Link 1 normalized initial capacity i/c x
Offered load
1 1
λ be the link’s aggregate arrival rate, and μ be the departure Algorithm 2: Congestion Aware Routing (CAR-M).
rate for each connection. Let t = 0 be the decision time—the Input: request pair r
time at which a request arrives. All future arrivals are assumed Input: current used capacity Ue , ∀e ∈ E
Input: total capacity Ce , ∀e ∈ E
to be accepted if capacity is available. Output: acceptance or rejection
If the link is loaded with i wavelengths at t = 0, where 1 foreach available shortest path p found for a connection generated
i < c, the expected number of future blocking in an arbitrary by request pair r by ASPF routing do
t 2 Compute the sum of congestion ratio;
time interval (0, t) can be written as 0 Pc,i (s)λds, where 3
Pc,i (s) is the probability that this link is full at time s. 1 Ce , CUe
γp = dUe e (2)
As t → ∞, the expected blocking diverges. To make a |p| e∈p
decision, we calculate the difference in blocking between
/* Path selection metric varies across
accepting and rejecting the request. We have two possible routing algorithms */
initial states: if we accept the request, the initial capacity 4 end
is i + 1. If we reject it, the initial capacity is i. The ex- 5 Path q ← p with min γp ; /* Optimality condition */
6 if |q| = T SLr then /* Always accept a Topological
pected difference in blocking (the opportunity cost)
is thus Shortest Path (TSP)
t t */
dci = limt→∞ 0 Pc,i+1 (s)λds − 0 Pc,i (s)λds . Since we 7 Accept the connection
8 end
assume that future arrivals are always accepted if there is free if γq > η then /* Threshold comparison
9 */
capacity, the difference Pc,i+1 (s)−Pc,i (s) decreases exponen- 10 Reject the connection;
tially, and the opportunity cost dci converges as t → ∞. 11 else
12 Accept the connection on route q;
We can use a continuous Markov chain to compute the 13 end
expected result for the opportunity cost for one link,
i i!
k=0 (i−k)!(cl)k
dc,l
i = c c!
(1) What happens when the traffic on two links is not indepen-
k=0 (c−k)!(cl)k dent? In practice, traffic dependence is difficult to model: the
flow pattern on the links depends not only on the expected
where the offered network load l is given by λ/μc. traffic demands but also on use as alternative routes. The
The opportunity cost for a single link is always less than complexity of the Markov chain for dependent links grows
one, thus we need to extend the model. For two independent quickly, so we use simulation. The right graph in Figure 3
links, the opportunity cost is equivalent to the sum of dc,l i shows results for various levels of dependence at offered load
for each one-link mode. Figure 2 shows a three-dimensional l = 0.9. Traffic dependence dp is the fraction of two-link
plot with the x- and y-axes representing the normalized initial traffic amongst all arrivals. When dp = 1, all connections use
capacity of each link. The z-axis is the opportunity cost both links, and the model reduces to a one-link model. When
computed as the sum of two one-link models. The threshold dp = 0, only the initial request uses both links, and the model
curve is shown in Figure 2 on the z = 1 plane. Above this is equivalent to the two-link independent model. The optimal
curve, a request should be rejected. The left graph in Figure 3 threshold curves drift upwards in x and y as dp increases,
shows several threshold curves (legends “mod ld=l”), which becoming a point at (1, 1) when dp = 1. The implication for
represent the ideal threshold at load l. Lower loads produce CAR-M is that the threshold value should be higher when a
higher thresholds and allow aggressive admission. At high path being considered has links with correlated congestion.
loads, few resources are available for long paths. If all links in a path of length L are independent, we might
The two-link model inspires our routing algorithm, Con- generalize CAR-M using η = 1/L. Instead, since links along
gestion Aware Routing based on Models, or CAR-M (see a path are often inter-dependent, we use a threshold of η = 0.5
Algorithm 2). For each shortest path, CAR-M computes the in all cases, which works well in practice and is robust both
average opportunity cost dc,l
i on each link using estimated load to topology and capacity scaling.
l = Ue /Ce , capacity c = Ce , and initial capacity i = Ue .
Routing algorithms must choose carefully among all shortest
IV. R ESULTS AND D ISCUSSION
available paths [17]. CAR-M does so by choosing the path
with the least average per-link opportunity cost. Paths of length This section provides simulation results. Figures 4-5 illus-
TSL are always accepted to avoid bias against node pairs with trate the properties of CAR-M and draw comparisons with
long TSLs. Longer paths are accepted only when the average related work. ASPF selects a shortest available path. WSP se-
opportunity cost is below the optimized simulation threshold lects a path with maximum residual capacity on the bottleneck
η = 0.5, which corresponds to the CAR-M curve in Figure 3. link. Neither algorithm is competitive with CAR-M, although
Since computation of γp takes total capacity into account, the the latter takes only ≈ 1.5× than ASPF to find routes.
threshold is adaptive to capacity scaling. Each ideal threshold The figures also compare CAR-M to two previous ap-
curve in Figure 3 is optimal for a homogeneous global traffic proaches. State-dependent routing (SDR)[27] approximates an
load l. At routing time, we know the operating point on the optimal solution using a Markov decision process and a link
plane, but not the offered load l. The threshold used at routing cost model. We also consider a throughput-competitive online
time must thus work for all load values, and η = 0.5 is routing and admission control algorithm (COL) [28]. The
analytically optimal for two independent links. original required knowledge of connection hold times. A later
5
0.1
[5] M. D. Feuer, D. C. Kilper, and S. L. Woodward, Optical Fiber Telecom-
munications. London, UK: Elsevier Inc., 2008, vol. B: Systems and
Networks, ch. ROADMs and their system applications.
Blocking probability [6] L. Berger, “Generalized Multi-Protocol Label Switching (GMPLS) Sig-
naling Functional Description,” RFC 3471, Jan. 2003.
0.01
[7] N. Ghani and S. Dixit, “Channel provisioning for higher-layer protocols
in WDM networks,” All-Optical Networking 1999: Architecture, Control,
and Management Issues, vol. 3843, no. 1, pp. 22–32, 1999.
[8] K. Casier, S. Verbrugge, D. Colle, I. Lievens, A. Groebbens, M. Pick-
ASPF
WSP avet, and P. Demeester, “Dimensioning studies for transparent optical
COL η=1 μ=1012
SDR η=1 backbone networks,” Transparent Optical Networks, 2005, Proc. of 7th
CAR-M η=0.50
0.001 International Conference, vol. 1, pp. 252–255 Vol. 1, July 2005.
0.75 0.8 0.85 0.9 0.95 [9] A. Elwalid, D. Mitra, I. Saniee, and I. Widjaja, “Routing and protection
Offered load
in GMPLS networks: From shortest paths to optimized designs,” IEEE
Journal on Lightwave Technology, vol. 21, no. 11, pp. 2828–2838, 2003.
Fig. 4. Routing algorithms on ARPANET with average link capacity 60. [10] Y. Xin, L. Battestilli, and G. Karmous-Edwards, “Generic network
services to support emerging grid applications,” in IEEE International
0.1 Conference on Broadband Communication, Networks, and Systems
ASPF
WSP (BROADNETS), Raleigh, North Carolina, USA, 9 2007.
COL η=1 μ=1050 [11] S. Spadaro, “Traffic engineering in IP over optical transport networks
SDR η=0.5
CAR-M η=0.50 for metropolitan and wide area environments,” Ph.D. dissertation, Uni-
versitat Politècnica de Catalunya, Dec. 2004.
Blocking probability