QoT-Aware Grooming, Routing, and Wavelength Assignment

China Communications January 2013 17

SELECTED PAPERS FROM IEEE ICCC'12

QoT-Aware Grooming, Routing, and Wavelength Assignment (GRWA) for Mixed-Line-Rate Translucent Optical Networks ZHAO Juzi1, Suresh Subramaniam1, Maïté Brandt-Pearce2

1Department of Electrical and Computer Science, The George Washington University, Washington DC 20052, USA 2Charles L. Brown Department of Electrical and Computer Engineering, University of Virginia, Charlottesville, Virginia 22904, USA

Abstract: A Mixed Line Rate (MLR) optical

network is a good candidate for a core back-

bone network because of its ability to provide

diverse line rates to effectively accommodate

traffic demands with heterogeneous bandwidth

requirements. Because of the deleterious ef-

fects of physical impairments, there is a

maximum transmission reach for optical sig-

nals before they have to be regenerated. Being

expensive devices, regenerators are expected

to be sparsely located and used in such a net-

work, called a translucent optical network. In

this paper, we consider the Grooming, Routing,

and Wavelength Assignment (GRWA) prob-

lem so that the Quality of Transmission (QoT)

for connections is satisfied, and the net-

work-level performance metric of blocking

probability is minimized. Cross-layer heuris-

tics to effectively allocate the sparse regen-

erators in MLR networks are developed, and

extensive simulation results are presented to

demonstrate their effectiveness.

Key words: MLR optical networks; transmis-

sion reach; cross-layer RWA; QoT-awareness;

regenerators

I. INTRODUCTION

Wavelength Division Multiplexing (WDM)-

based optical networks are widely deployed at

the Internet’s core because of their ability to

carry large amounts of traffic. A Mixed-Line-

Rate (MLR) optical network (with 10/40/100

Gb/s rates per wavelength) is a good candidate

for next generation core backbone, because

the multiple line rates are more suitable for

heterogeneous traffic requirements; e.g., very

high bit-rate demands can be carried by 100

Gb/s lightpaths, while lower bit-rate traffic

connections can use 10 or 40 Gb/s lightpaths.

Optical networks present the challenges of

wavelength continuity and physical layer im-

pairments. Wavelength continuity mandates

that a single wavelength be assigned on all the

links of a transparent segment (i.e., between

regeneration points) of a lightpath. Due to

physical impairments, it is impossible to con-

struct purely optical core networks. Regen-

erators (called 3R regenerators because of

their function of re-amplification, retiming,

reshaping) can clean up the accumulated im-

pairments by performing an Optical-Electrical-

Optical (OEO) conversion and processing the

electrical signal. Being expensive devices,

these regenerators are expected to be sparsely

located in the network, and networks with

sparse regeneration capabilities are called

translucent optical networks.

In physically-impaired networks, connec-

tions have Quality of Transmission (QoT)

Received: 2012-10-18 Revised: 2012-12-07 Editor: NIU Zhisheng

18 China Communications January 2013

requirements (as exemplified by, say, their Bit

Error Rates (BER)), which must be satisfied

by the network. At the same time, network

operators are interested in utilizing the net-

work efficiently, or in other words, minimiz-

ing the blocking probability of connections. In

this paper, crosslayer heuristics to efficiently

groom and route dynamic multibit-rate con-

nections and allocate the OEO converters in

regenerators are developed for MLR optical

networks. Extensive simulation results are

presented to demonstrate their effectiveness in

reducing the blocking probability while en-

suring connections’ QoT. These algorithms are

QoT-aware (or Impairment-Aware, IA) in the

sense that they consider the connections’ QoT

while selecting the routes.

The existing work on routing in MLR opti-

cal networks is summarized below. Following

that, we highlight this paper’s contributions.

1.1 Related work

Most of the work with MLR deals with static

traffic, i.e., all the demands are known a priori.

The authors of Refs. [1-4] consider QoT-aware

routing with the objective to minimize the cost

of transponders. Various Integer Linear Pro-

gramming (ILP) formulations and effective

heuristics are proposed. The authors of Ref. [5]

design a QoT-aware cost-effective transparent

MLR network that provides dedicated protec-

tion at the lightpath level. ILP and heuristics

for an energy-efficient MLR optical network

are proposed in Refs. [6-7]. In Ref. [8], the

authors study reliable and cost-effective

QoT-aware routing. An IA Routing and Wave-

length Assignment (IA-RWA) algorithm called

IA-KS-EDP was proposed in Ref. [9]; it con-

siders the effects of several impairments and

uses edge disjoint paths to satisfy static traffic

demands.

Some research on dynamic traffic has been

done recently. In Ref. [10], the authors present

a QoT (in terms of BER) model considering

Cross-Phase Modulation (XPM). In addition,

novel dynamic lightpath provisioning in sev-

eral MLR scenarios is investigated. The prob-

lem of IA-RWA in MLR Wavelength-Swit-

ched Optical Networks (WSONs) with two

Path Computation Element (PCE) architec-

tures is addressed in Ref. [11], and a novel

PCE protocol extension is introduced. How-

ever, these two papers consider transparent

networks without regenerators. The authors of

Ref. [12] propose heuristics for dynamic con-

nection RWA in MLR optical networks, but

QoT is not considered. A recent paper that

considers grooming and IA-RWA for dynamic

traffic is Ref. [13]; however, the network

model used in Ref. [13] is different from what

is used in this paper.

1.2 Contributions

As discussed, most of the work in MLR net-

works has been for static traffic. Most of the

existing work on dynamic traffic considers

either transparent networks without regenera-

tors, or completely ignores impairments. This

paper addresses the problem of dynamic QoT-

aware Grooming, Routing, and Wavelength

Assignment (GRWA) in MLR translucent op-

tical networks. We present a QoT-aware algo-

rithm for GRWA (called Logical Graph, LG)

that effectively allocates the sparsely available

OEO converters to connections, so that the

blocking probability is reduced. Then we pre-

sent comparisons of our algorithm with other

algorithms that are either simpler or are

QoT-unaware to demonstrate the advantages

of our algorithm.

The paper is organized as follows. In Sec-

tion II, we present the network model in our

study. Section III describes the QoT-aware

RWA algorithms. Section IV presents and

discusses the simulation results. We conclude

the paper and present possible extensions to

this work in Section V.

II. NETWORK MODEL

The network is modeled as a graph G(V, E),

where V is the set of nodes in the network,

|V|N = , and E is the set of links. Each link

e EÎ consists of two fibers with W wave-

lengths each, oriented in opposite directions.

Our algorithm (called Logical Graph) for QoT- aware grooming, rou-ting, and wavelength assignment in Mixed- Line-Rate networks shows good perform-ance.


Bidirectional connection requests with 10 G,

40 G, or 100 G bandwidth requirement be-

tween a pair of nodes (s, d) randomly arrive to

and depart from the network1. The objective is

to assign a route and wavelength(s) and allo-

cate OEO regenerators as needed to minimize

blocking.

2.1 MLR model

We assume that the line rates in the network

are 10 G, 40 G, or 100 G with different modu-

lation methods. Each wavelength on any link

can be modulated at any one of these line rates

with the corresponding modulation scheme, as

assumed in Refs. [4, 10], as long as the corre-

sponding transponder is available.

A transponder is used to initiate and termi-

nate lightpaths. It is assumed to function at a

specific line rate, e.g., a 40 G transponder can

only be used for launching a 40 G lightpath

and not a 10 G or 100 G lightpath. Each node

in the network has a limited number of trans-

ponders for each line rate. However, we as-

sume that a transponder at a node can be

shared across all links at the node (share-per-

node model).

2.2 3R model

A 3R node uses one of a limited number of

OEO converters at the node to regenerate an

optical signal. Like transponders, OEO con-

verters are also line-rate-specific, and are

available for use by any connection (at the

specific line rate) traversing the node. An

OEO converter can also provide wavelength

conversion, if necessary. For each line rate,

there is a set of 3R nodes in the network. The

set of nodes for one line rate may be different

from the set for another rate, and a single node

may be a 3R node for more than one line rate.

We introduce a few definitions here. A 3R

node will be called an OEO node if it has at

least one OEO converter available for use, i.e.,

if not all OEO converters are being used by

ongoing connections. An allocated OEO node

is an OEO node providing OEO converters to

a connection. Further, the transparent path

between two OEO nodes is called an OEO

segment. In this definition, the intermediate

nodes may have OEO converters, but they are

not allocated to the path in question. A con-

secutive OEO segment on a path is the seg-

ment between two consecutive OEO nodes

(whether they are allocated to the connection

or not).

2.3 QoT model

We assume that a connection’s QoT require-

ment can be met as long as there is no segment

that is longer than the transmission reach (TR)

(as in Refs. [4, 6, 8]). For each line rate, there

is a specific TR that ensures that the QoT of a

connection at that line rate is satisfactory. If

the length of a transparent segment between

two allocated OEOs is not greater than the

corresponding TR, the QoT of the segment is

acceptable; otherwise, it cannot be a subpath

for a connection.

2.4 Grooming model

We assume that grooming can be done only at

the end nodes of a connection, and not at in-

termediate nodes. For example, two 10 G

connections between the same source and des-

tination may be groomed onto a 40 G lightpath

between those nodes, but a connection that

does not have the same source and destination

may not be groomed onto this lightpath. This

simplifies network design, because it does not

require the 3R nodes to perform grooming in

addition to regeneration.2

2.5 Splitting model

Similar to the grooming model, the end nodes

of a connection can split (at source) and com-

bine (at destination) multiple paths, which are

used to satisfy that connection’s bandwidth

demand. For example, a 100 G connection

may be split and satisfied using two 40 G and

two 10 G lightpaths between the same source

and destination. Whether a connection is as-

signed a single path or multiple paths depends

on the algorithm policy and the network state

(e.g., available network resources) when the

1. The focus of this paper is on RWA for MLR net-works. As such, we as-sume that bandwidth requirements of sub- wavelength connections are aggregated so that only 10 G, 40 G, or 100 G requests are made to the network. 2. It may be possible to improve the performance with multi-hop grooming, as in Ref. [17], but we do not consider that in this paper. Note that the per-formance improvement is not guaranteed because of the extra transponders that would be used for grooming at intermediate nodes.


connection arrives.

2.6 Performance metrics

Since the bit rate (bandwidth) requirements of

connections are different, instead of the usual

connection blocking ratio, two metrics related

to bit rate of connections are used. Let iB

denote the bandwidth requirement of connec-

tion i, and it denote the duration of the con-

nection (i.e., the time from setup to termina-

tion of the connection). Further, let J be the set

of connections that arrive to the network, and I

be the set of blocked connections ( I JÍ ). We

define:

1) Bandwidth blocking ratio (BW Blocking)

BWBlockingi

i I

jj J

B

BÎ

Î

=å

å

2) Throughput blocking ratio (TH Blocking)

THBlocking .i i

i I

j jj J

B t

B tÎ

Î

=å

å

A connection can be blocked due to any of

three reasons: (a) there are not enough trans-

ponders for the connection, (b) lack of wave-

length, or (c) the algorithm cannot find a path

with every segment shorter than the TR. Calls

are typically blocked for a combination of

reasons. Although multiple paths may be used

for a connection, a connection is accepted into

the network only if its bandwidth requirement

can be fully satisfied. Connections whose

bandwidth demands can only be partially sat-

isfied are blocked. The objective is to select

paths and effectively allocate the OEO con-

verters to the connections so that the reach

limit is met and the blocking probability is

minimized.

III. ALGORITHMS

We now present our GRWA algorithms. Our main algorithm is called LG as the algorithm is based on constructing a logical graph. We also present an algorithm called SP that is

based on shortest physical length paths, and a couple of other variants for benchmarking purposes. Suppose a new connection request (with bandwidth 10 G, 40 G, or 100 G) arrives between source s and destination d.

Since the existing (i.e., already established)

lightpaths between s and d may have some

residual capacity available, the first step in all

of the algorithms is to groom as much of the

bandwidth demand of the new connection into

those lightpaths. Note that grooming onto an

existing lightpath does not need any new re-

sources (transponders, OEOs, wavelengths) to

be allocated. The algorithms may differ in

which lightpaths are selected for grooming, if

many are available. If the grooming step can

completely satisfy the bandwidth demand of

the new connection, the algorithm terminates.

After the grooming step, the remaining re-

quested bandwidth can be any multiple of 10

Gb/s, up to and including 100 Gb/s. There

may exist many potential combinations of

different newly created paths of different

line-rates for satisfying the remaining band-

width. All the possible combinations for each

value of remaining bandwidth are shown in

Figure 1.

3.1 Logical Graph (LG)

Our algorithm considers the availability of

transponders, OEOs, and wavelengths as well

as the TR when providing path(s) to a connec-

tion. Before describing the algorithm, some

notation is needed. The number of currently

available transponders with line rate r at node

n is denoted as ( )rT n ; the number of trans-

ponders allocated to a path p is ( )prt n ; the

number of currently available number of

OEOs is ( )rO n ; the number of OEOs allo-

cated to a path p is ( )pro n .

Grooming Step: recall that only existing

paths with same source and destination are

considered. If a lightpath’s residual capacity is

greater than the requested bandwidth of the

new connection, we simply set it equal to the

requested bandwidth. Then, we assign a cost

to each existing lightpath as follows:


Fig.1 Path line-rate combinations for different remaining bandwidth requests after grooming

1 1 1

max , ,max( ) ( ) ( )n

r r rT s T d O n

ì üï ïï ïí ýï ïï ïî þ , where n is

any allocated OEO node on the existing path,

and r is the line rate with the existing path.

Existing paths are sorted in decreasing order

of residual capacity, and the paths with the

same residual capacity are sorted in increasing

order of this cost. The idea behind assigning

costs in this manner is to ensure that a path

and line-rate with many available transponders

is chosen for the connection, thereby leaving

more resources for future connections. By

following the sorted order of paths, as much of

the new connection’s demand as possible is

satisfied using existing paths. The time com-

plexity for the grooming step can be shown to

be ( log )O Hq H H+ , where H denotes the

maximum number of existing lightpaths in the

network; and q denotes the number of 3R

nodes in the network.

Remaining Bandwidth Step: for the remain-

ing bandwidth, one or more logical graphs will

be created sequentially if possible – one logi-

cal graph ( , )rG V E¢ ¢ ¢ for each required path

with line rate r (if the corresponding trans-

ponders at the source and destination are

available). For example, for the combination

2*40G + 2*10G paths for 100 G remaining

bandwidth, two logical graphs with r = 40 and

two logical graphs with r = 10 will be created

sequentially if both the source and destination

have at least two 40 G transponders and two

10 G transponders.

We next explain how a logical graph for a

particular rate r is constructed. V' includes the

OEO nodes at line rate r as well as the source

s and destination d of the connection. A logical

link ,x ye¢ exists between two logical nodes x

and y if there exists at least one path in the

physical graph G whose length is no more

than the TR for r and the path has at least one

wavelength available. The logical link ,x ye¢ is

represented by two directed edges, one from x

to y, with cost set as 1

( )rO y (if y = s or y = d,

the cost is set to zero), and the other from y to

x, with cost set as 1

( )rO x (if x = s or x = d,

the cost is set to zero).


The physical path for a logical link ,x ye¢ is

selected as follows: K' (an algorithm parame-

ter) shortest paths are pre-computed [14] be-

tween x and y, and the path with most avail-

able wavelengths is selected; if there are many

such paths, then the one with minimum num-

ber of hops is selected.

After the logical graph is constructed, a

logical path from s to d with minimum cost is

found by Dijsktra’s shortest path algorithm.

An OEO is allocated at each node along the

logical path, and the concatenation of the

physical links forms the physical lightpath for

the request. The First-Fit (FF) wavelength is

assigned to each allocated OEO segment. The

time complexity to find a new candidate path

can be shown to be 2 3( ' )O K WNq q+ , where

W denotes the number of wavelengths per

fiber and N denotes the number of nodes in the

network.

We illustrate the logical graph construction

with the example in Figure 2. Suppose the

source and destination are nodes 0 and 5, re-

spectively, in the original (physical) graph,

shown in Figure 2(a), and let nodes 2 and 3 be

the OEO nodes. Suppose the logical links ob-

tained after wavelength availability and reach

are checked are as shown in Figure 2 (b).

Suppose the currently available number of

OEOs at node 3 is 5, and at node 2 is 4; then

the costs of the directed logical edges to node

3 and 2 are set to 1/5 and 1/4, respectively, and

the costs of directed logical edges to node 0

and 5 are set as 0. Running Dijsktra’s shortest

path algorithm produces the logical path 0-3-5 with cost 1/5. Then, this path is allocated to

Fig.2 Example showing (a) a physical topology and (b) the corresponding logical topology

the connection, and node 3 is the allocated

OEO node. The FF wavelengths on OEO

segments 0-3 and 3-5 are assigned.

Now, recall that the remaining bandwidth is

satisfied using one of the possible combina-

tions shown in Figure 1. The particular com-

bination to satisfy the remaining bandwidth is

selected as follows. A cost C is assigned for

each combination c. For a combination c with

path set cP , and for each allocated OEO node

n, let

( ) ( )c

c pr r

p P

o n o nÎ

= å

( ) ( )c

c pr r

p P

t s t sÎ

= å

and

( ) ( )c

c pr r

p P

t d t dÎ

= å

Then define

( ) ( ) ( )max max , ,

( ) ( ) ( )

c c cr r r

cr R n r r r

o n t s t dC

O n T s T dÎ

ì üï ïï ï= í ýï ïï ïî þ

where set R includes the line rates that are

involved in the combination (e.g., R = {10,40}

for the combination 2*40G+2*10G). The se-

lected combination is the one with minimum

cost. Roughly speaking, this cost function picks

a combination so that the number of allocated

transponders and OEOs is as small as possible

(relative to the number deployed at the nodes).

The complexity of the entire LG algorithm is 2 3( ' log )O K WNq q Hq H H+ + + .

3.2 Shortest Path (SP)

This algorithm gives preference to paths with

shorter physical length.

Grooming Step: existing paths are sorted in

decreasing order of residual capacity (if re-

sidual capacity is greater than bandwidth re-

quirement of the new connection, set residual

capacity as the bandwidth requirement), and

the paths with same residual capacity are

sorted in increasing order of their lengths. This

order is used to satisfy as much of the band-

width demand as possible for the new connec-

tion. The time complexity for the grooming

step is ( log )O H H . Remaining Bandwidth Step: for each possi-


ble combination for the remaining bandwidth,

new lightpaths of this combination are found one by one. When finding each new path, the

algorithm first selects a path and allocates

OEOs, and then assigns a wavelength to each allocated OEO segment. In the path selection

and OEO allocation step, the algorithm first

uses the K' shortest pre-computed paths as candidate paths (same path set used by the LG

algorithm for finding paths for logical links),

belonging to ,s dP for node pair s-d. Given a candidate path ,i s dp PÎ , OEO allocation on

ip is done as follows. Starting from node s,

find the furthest OEO node x with the property that: if node x' is the next OEO node after x,

there is at least one available wavelength on

OEO segment s-x and the length of s-x is not greater than the TR for the corresponding rate

with the path, but no available wavelength can

be found for the segment s-x' or the length of s-x' is longer than TR. This means that node x

needs to serve as a wavelength converter or

regenerator, i.e., node x


quire an exponential service time with unit

mean. The source and destination of the con-

nection are randomly picked. The bandwidth

requirement of the connection is 10 G, 40 G, or

100 G with probability 25%, 50%, and 25%,

respectively. For each data point in the graphs,

we simulated several instances between 10 000

and 100 000 connection arrivals, and obtained

95% confidence intervals.

We present results for the European Optical

Network (EON) and USA network (USANET)

(Figures 3 and 4) topologies. The 3R nodes

are selected according to the Greedy algorithm

presented in Ref. [15]; the algorithm is applied

to each line rate separately. The 3R node

placement in EON and USANET are shown in

Table I. The number of transponders per line

rate at each node is found according to Ref.

[16], which considers the expected load of the

network (i.e., the number of connections in the

network) and the probabilities of different line

rate connections. The number of transponders

with each line rate for EON and USANET are

shown in Table I. The TR values used in this

paper are from Ref. [6]: they are 1 750 km,

1 800 km and 900 km, respectively. Note that

the TR for 40 G is actually higher than for

10G because a different modulation method is

assumed [6]. We assume that K' = 40, i.e., up

to 40 shortest paths are computed offline for

each node pair. In addition, K = 3 alternate

paths are used by the algorithms. Each fiber is

assumed to have 32 wavelengths.

Fig.3 28-node EON. The number on each link corresponds to the length in km


Fig.4 24-node USA network (USANET). The number on each link corresponds to the length in km Table I 3R node placement and number of transponders

3R nodes Number of transponders per node Network

10 G 40 G 100 G 10 G 40 G 100 G

EON 5,19 10,13 9,10,12,17,19,25 9 15 9

USANET 9 9 7,12,21 10 17 10

4.1 Performance comparison of algorithms

Figures 5-8 show the BW Blocking and TH Blocking versus the network Erlang load for the various algorithms with 10 OEO convert-ers and 30 OEO converters per 3R node in EON, and with 16 OEO converters and 34 OEO converters per 3R node in USANET. The loads were selected so that blocking probabilities fall in the 410- to 110- range.

As can be seen, QoT-G, which ignores the TR (and therefore physical impairments) during RWA, performs much worse than the other algorithms that are QoT-aware. Interestingly, the blocking ratio stays essentially flat. This can be explained as follows. Recall that QoT-G allocates OEO converters to paths considering only wavelength availability and not the TR. When the selected path is tested to see if every OEO segment is shorter than TR, there is a high probability that the test fails, and hence

Fig.5 BW Blocking and TH Blocking vs. Erlang load; 10 OEO converters per 3R node in EON


Fig.6 BW Blocking and TH Blocking vs. Erlang load; 30 OEO converters per 3R node in EON

Fig.7 BW Blocking and TH Blocking vs. Erlang load; 16 OEO converters per 3R node in USANET

Fig.8 BW Blocking and TH Blocking vs. Erlang load; 34 OEO converters per 3R node in USANET the connection is blocked. When the load in-

creases, a wavelength-continuous path from

source to destination becomes harder to find,

and therefore, OEO converters are allocated to

the connection for doing wavelength conver-

sion. Since these OEO nodes also perform 3R

regeneration, the probability that the selected

path will be rejected by the test does not in-

crease.

Further, LG clearly outperforms the other


algorithms. The significant difference between

LG and LG-single path shows that the network performance can be greatly improved by multi-path provisioning (path splitting), par-ticularly at lower loads. Interestingly, LG-single path outperforms SP when the number of OEOs is larger, despite the fact the SP allows connection splitting. We explain this a little later. Both BW Blocking and TH Blocking have similar trends, indicating that these re-sults are not dependent on connection holding times, which are not assumed to be known when the connection is set up.

For the EON network, two more simula-tions are done with different numbers of OEO converters at different 3R nodes. Specifically, in the first simulation, for 10 Gb/s and 40 Gb/s,

the first 3R node has 10 OEO converters, and

the second has 6 OEO converters; for 100 Gb/s,

of the six 3R nodes, the first three have 10

OEO converters each, and the other three have

6 OEO converters each. In the second simula-

tion, for 10 Gb/s and 40 Gb/s, one 3R node

has 30 OEO converters, and the other has 20

OEO converters; for 100 Gb/s, the first three

3R nodes have 30 OEO converters each, and

the other three have 20 OEO converters each.

Figures 9-10 shows the results of BW Block-

ing and TH Blocking versus the network Er-

lang load for the various algorithms. These

results further confirm that the performance

improvement of our algorithm is not sensitive

to parameter changes.

Fig.9 BW Blocking and TH Blocking vs. Erlang load; 10/6 OEO converters per 3R node in EON

Fig.10 BW Blocking and TH Blocking vs. Erlang load; 30/20 OEO converters per 3R node in EON


4.2 Blocking vs. number of OEO converters

The blocking ratios are plotted as a function of

the number of OEO converters per 3R node

for EON in Figure 11 and for USANET in

Figure 12. Once again, the superior perform-

ance of LG is obvious. As we can see, there is

a significant decrease in blocking as the num-

ber of converters increases for both LG and

LG-single path, pointing to the ability of these

algorithms to wisely allocate OEO converters.

An interesting observation is that when the number of OEO converters is small, the

performance of LG-single path is worse than SP’s; but LG-single path outperforms SP de-spite no connection splitting, when the number of OEO converters is large. SP prefers shorter length paths, but these paths end up using more OEOs as wavelength converters; in other words, longer paths that need fewer OEO nodes may not be selected by SP. But for LG-single path, the OEO nodes are used when they are needed as wavelength converters or regenerators, thus they are used more effi-ciently. On the other hand, when the number of OEOs is small, the single path restriction dominates and SP performs better.

Fig.11 BW Blocking and TH Blocking vs. OEO converters per 3R node in EON; load = 130 Erlangs

Fig.12 BW Blocking and TH Blocking vs. OEO converters per 3R node in USANET; load = 100 Erlangs


erator Location Problem[J]. Networks, 2010, 55(3): 205-220.

[16] RIVAL O, MOREA A. Resource Requirements in Mixed-Line-Rate and Elastic Dynamic Optical Networks[C]// Proceedings of the 2012 Opti-cal Fiber Communication Conference and Exposition/National Fiber OpticEngineers Conference (OFC/NFOEC): March 4-8, 2012, Los Angeles, USA, 2012.

[17] SHEN Gangxiang, TUCKER R S. Sparse Traffic Grooming in Translucent Optical Networks[J]. Journal of Lightwave Technology, 2009, 27(20): 4471-4479.

Biographies ZHAO Juzi, is a Ph.D. candidate in the Department of Electrical and Computer Engineering at the George Washington University, USA. She received an M.S. degree from the same university in 2009. Her current research is in optical and data center networks.

Suresh Subramaniam, is a Professor of Electrical and Computer Engineering at George Washington Uni-versity, USA. He received the Ph.D. degree in electrical engineering in 1997 from the University of Washing-ton, Seattle, USA. His research interests include opti-cal, wireless and datacenter networks, and he has published over 140 refereed articles in these areas. He serves on the editorial boards of several journals and is TPC Co-Chair of INFOCOM 2013. He is a Senior Member of the IEEE. Maïté Brandt-Pearce, received her Ph.D. in electrical engineering in 1993 from Rice University, USA. She is currently a full professor in the Department of Elec-trical and Computer Engineering at the University of Virginia, USA. Her research interests include optical and wireless communications, and biomedical and radar signal processing. She is a member of Tau Beta Pi, Eta Kappa Nu, and a senior member of the IEEE. Dr. Brandt-Pearce has over a hundred major publications.

Documents

QoT-Aware Grooming, Routing, and Wavelength Assignment