You are on page 1of 8

Optical Packet Switching: A Network Perspective

Franco Callegati, Walter Cerroni, Giorgio Corazza, Carla Raffaelli


D.E.I.S. - University of Bologna
Viale Risorgimento, 2 - 40136 Bologna - ITALY
{fcallegati,wcerroni,gcorazza,craffaelli}@deis.unibo.it

Abstract In the last few years research on WDM optical


packet-switching has mainly concentrated on issues at the single
node level. The goal of this paper is to discuss about the problems
arising when the focus is moved toward a network-wide scope. In
particular, an overview of routing techniques that may be adopted
in an optical packet-switched backbone is presented, showing
the effects of adaptive multi-path routing strategies on the
network performance. Furthermore, performance differentiation
based on different routing and contention resolution strategies
is proposed and analyzed in a simple QoS-aware scenario.
Then, the application of adaptive routing strategies to network
recovery in case of single link failure and the effects of dynamic
multi-wavelength management and multi-path routing on packet
sequence are also discussed.

I. I NTRODUCTION
The advances experienced in the last decade by photonic
technology have made optical networking a very good candidate to implement the very high-capacity backbone of future
communication networks. The WDM optical circuit-switching
paradigm (at either fiber or wavelength level) is a technique
to realize such optical backbone with some flexibility in terms
of resource provisioning, and offers huge bandwidth capacity
to the end-user. Nonetheless, this approach provides access to
bandwidth with a very coarse granularity and therefore with
limited QoS management capability.
Optical Burst Switching (OBS) [1] and Optical Packet
Switching (OPS) [2] are respectively a medium and a longer
term networking solution, promising more flexibility and
efficiency in bandwidth usage combined with the ability
to support diverse services [3]. Research activities on OPS
have mainly concentrated on issues at the single node level,
studying architectural and algorithmic solutions to provide
satisfying levels of performance [4]. In particular, the problem
of congestion resolution has been extensively studied, showing
that smart policies able to effectively exploit both the time and
the wavelength domains can make a huge difference in terms
of average packet loss and latency [5].
The aim of this paper is to introduce and discuss about
some of the problems arising when the focus is moved from
the single node perspective toward a network-wide scope,
with an overview of adaptive, multi-path routing techniques
that may be adopted in an optical packet-switched backbone.
Also different routing and contention resolution strategies for
service differentiation can be proposed and analyzed in a QoSaware scenario based on the DiffServ model [6]. However, the
number of QoS classes must be kept as small as possible in order to minimize operational efforts, since complex scheduling

algorithms may not be applicable because of the peculiarity


of queuing in the optical domain [7].
Multi-path routing strategies may also be exploited in order
to provide reliability to the optical packet-switched network. In
particular, when a single link failure occurs, packets previously
routed on that link may be transmitted on alternative paths,
according to the multi-path routing strategy adopted. Additionally, an important issue that should be taken into account is
the impact of wavelength management and multi-path routing
strategies on the packet sequence, which can be easily broken,
for instance, when the difference in terms of latency along
different routing paths is not negligible.
The paper is structured as follows: first, the behavior assumed at the single node level is addressed in section II;
then, an overview of multi-path, adaptive routing techniques
that may be deployed in an OPS network is presented in
section III, with particular attention to traffic differentiation
in section IV; the application of adaptive routing strategies to
network recovery in case of single link failure is discussed
in section V, while a methodology for the evaluation of the
effects of dynamic wavelength and routing management on
packet sequence is introduced in section VI; finally, section
VII concludes the work.
II. S INGLE NODE BEHAVIOR
The reference scenario considers a packet-switched optical
network switching asynchronous, variable-length packets, statistically multiplexed over multi-wavelength links. A general
switching node with N input and N output fibers, carrying
W wavelengths each, is considered. Each node is equipped
with Fiber Delay Line (FDL) buffers, which are able to delay
optical packets for an amount of time multiple of a given time
unit D (called buffer granularity). The buffer size is related to
the maximum available delay DM AX which depends on the
number B of delay lines used.
Contentions due to contemporary packet arrivals at a given
output port are resolved by means of load balancing techniques
that exploit both the wavelength and time dimensions. In
general, after the output fiber has been determined for a
given packet by the routing strategy, the switch control logic
must face a two-dimensional scheduling problem: the choice
of the wavelength to transmit the packet on and, in case
of contention, the delay to be assigned to that packet. This
problem can be referred to as the Wavelength and Delay
Selection (WDS) problem [5].

Two classes of WDS algorithms, characterized by a very


similar computational complexity, are the following:
delay oriented algorithms (D-type), that aim at minimizing the waiting time of a queued packet and therefore
act according to the principle that, when a packet has to
be queued, it will join the shortest available queue (the
shortest delay provided by the FDL buffer);
gap oriented algorithms (G-type), that aim at minimizing
the gaps (caused by the discrete number of delays available) between packets and, consequently, maximizing the
throughput of the switching matrix, acting according to
the principle that, when a packet has to be queued, it will
be sent to the delay that is closest to the transmission end
of the preceding packet.

G-type

1
2
3
4

D-type
t0
Fig. 1.

t0+D

Different node-level packet scheduling policies

Packet Loss Probability

1.0e-1
1.0e-2

1
2
3
4

D-type

Fig. 3.

G-type

1.0e-5

t0

t0+D t0+2D t0+3D t0+4D

t0

t0+D t0+2D t0+3D t0+4D

1
2
3
4

1.0e-3
1.0e-4

LP
HP

t0+2D t0+3D t0+4D

The two approaches are briefly sketched in Fig. 1, for the


case of an output fiber with W = 4 wavelengths and B = 4
delays (D, 2D, 3D, 4D). Other policies based on the voidfilling principle proposed in [8] are not considered here due
to their higher computational complexity.
A G-type buffer policy will be assumed in the following,
since it proves to realize a good trade-off between complexity
and performance. In fact, this choice results in a lower packet
loss rate compared to a D-type wavelength assignment policy,
as shown in Fig. 2 for a reference case with N = 4, W =
16, B = 7.

When considering traffic differentiation, incoming packets


are assumed to belong to two classes, namely High Priority
(HP) and Low Priority (LP) optical packets. Since FDL
buffers cannot allow a full random access to the queue [7],
conventional preemptive or priority queuing techniques are not
easily applicable here. Thus QoS management must mainly
rely upon mechanisms based on differentiated scheduling and
a-priori access control to the optical buffer, by adopting, for
instance, either a time-based or a wavelength-based resource
partitioning technique [9].
Considering the limitations of todays optical technology, a
partitioning policy relying on the wavelength domain may be
more effective and provide more flexibility than a partitioning
policy exploiting the time domain. In the following, a wavelength is considered congested when the corresponding FDL
buffer is full and no more packets are allowed into it. Under
this assumption, the shared wavelength partitioning strategy
reserves any K wavelengths to HP traffic based on the actual
buffer occupancy; namely when more than K wavelengths are
not congested, both LP and HP packets may be transmitted,
otherwise when only K or less wavelengths are congested
(whichever they are), only HP packets can be transmitted.
Figure 3 shows an example for K = 2, W = 4, B = 4.

LP
HP

Wavelength-based partitioning policy for QoS differentiation

III. A DAPTIVE ROUTING IN OPS NETWORKS


1.0e-6

0.5

1.5

2.5

3.5

D (normalized to the average packet length)


Fig. 2.

Performance of different node-level packet scheduling policies

Generally speaking, routing algorithms can be either static


or adaptive, i.e. dynamic. The former define static routing
tables once and for all, whereas the latter route traffic by
exploiting information regarding the state of the network. The

1e-03

Packet Loss Probability

adaptive algorithms may be further specialized depending on


the number and cost of the paths that are considered in order
to take the routing decision [10]. This paper assumes a meshed
network topology with WDM links.
The basic idea is to combine the flexibility of adaptive
routing with the efficiency of packet multiplexing over a large
set of wavelengths by means of an effective WDS policy. At
each node, traffic is normally forwarded along the shortest
path but alternative paths of equal or higher hop count are also
identified and may be used. Therefore, two possible routing
strategies may be defined [11]:
Shortest-Path Routing (SPR), based on minimum hop
count and not using any alternative path, i.e. a static
choice;
Multi-Path Routing (MPR), including alternative paths
that are dynamically used by the network nodes when
the link along the shortest path (also called the default
link) becomes congested.
The first decision to be taken concerns how many alternative
paths, among those possible, should be taken into account
for load balancing in MPR. Once the alternative routes have
been defined, each node tries to send the incoming packets
on the default link and, in case this one is congested, i.e.
there is no wavelength available and the buffer is full, one of
the alternative paths is used, picking it up from one or more
alternative sets. In case also the alternative paths do not have
available wavelengths and/or delays, the packet is dropped.
The following options may be applied in the definition of the
alternative sets of routes:
no alternative path: this is the case of SPR;
shortest alternative paths (SAP): beside the default link
an alternative set of routes is considered, which includes
any other shortest path different from the default one, i.e.
any other path with the same hop count as the shortest
one; depending on the network connectivity, such an
alternative set may or may not exist;
n-shortest alternative paths (n-SAP): besides the default
link, n alternative sets of routes are considered, where the
i-th set includes every path with i 1 hops more than
the shortest one; obviously, the first set (i = 1) does not
include the default shortest path itself;
An additional, more dynamic approach to the load balancing
policy consists into applying the WDS policy not on a single
link (either default or alternative) but on an entire set of links,
which can be defined as follows:
shared shortest paths (SSP): WDS is performed directly
taking into account all the wavelengths on any shortest
path link, including the default one;
n-shared shortest paths (n-SSP): the WDS is performed
directly taking into account all the wavelengths on any
link belonging to paths with up to n 1 hops more than
the shortest one.
A comparison of the behavior of the aforementioned MPR
strategies in terms of average packet loss rate is shown in
Fig. 4, evaluated over the European optical network topology

1e-04

1e-05

1e-06

SPR
Fig. 4.

SAP

2-SAP

SSP

2-SSP

Performance of different MPR strategies

discussed in [11]. The figure shows that a MPR approach starts


to be effective when also non-shortest paths are taken into
account (only the case for paths with up to a single hop more
than the shortest one have been considered).
IV. Q O S SUPPORT
When the incoming traffic is classified into priority classes,
QoS differentiation can be achieved by differentiating the
concept of congestion and/or providing different alternatives
to LP and HP traffic. As an example, congestion may be
defined according to the wavelength availability on the default
link, adopting a strategy similar to the shared wavelength
partitioning used at the single node level. The value of K
may be different for different classes of service, giving the
capability to decide which class of traffic makes use of the
alternative paths more frequently. However, the use of nonshortest, alternative paths may cause the packets to stay longer
in the network and the transmission delay to become much
higher than the SPR case. Therefore, in the following it is
assumed that MPR strategy may be used for LP traffic only,
while HP traffic is always routed according to the SPR strategy,
aiming at preserving the HP traffic stream as intact as possible.
An evaluation of such strategies over a simple 5-nodes
meshed topology, under the assumptions considered in [12],
is presented in Fig. 5. The plot shows the good level of
differentiation between losses of HP traffic (20% of the total)
and LP traffic (80%), when only the LP traffic is re-routed
on alternative paths in case of congestion on the default link
with a 2-SAP strategy (curves labeled MPR) or when the MPR
in not applied at all (curves labeled SPR). As expected, the
higher the number of reserved wavelengths K, the higher the
gain obtained by the HP traffic, while the performance of
the LP traffic is barely affected in the range considered here.
Furthermore, HP traffic is also not affected by the different
routing strategies applied to LP packets.
On the other hand, according to Fig. 6 where K was set
to 3, for a very low percentage of HP traffic a good level of
performance may be achieved. Of course, when the percentage
of HP traffic grows over the 20%, the service differentiation
strategy becomes less effective, although the LP traffic seems

Packet loss probability

1e-02

SPR LP

1e-04
1e-05

MPR LP

10

11

12

13

14

15

SPR HP

1e-06
1e-07
1e-08

Fig. 7.

MPR HP
1

2
3
4
No. of reserved wavelenghts (K)

Fig. 5. Performance of QoS differentiation by means of routing and resource


partitioning techniques

1
Packet loss probability

1e-01

1e-03

1e-01
1e-02

MPR LP

1e-04

MPR HP

1e-05

SPR HP

1e-06
1e-07
1e-08

Network topology with average node degree E = 3.25


0

10

11

12

13

14

15

SPR LP

1e-03

10

20
30
40
Percentage of HP traffic

50

Fig. 6. Performance of QoS differentiation by means of routing and resource


partitioning techniques

to be still slightly affected. It follows that in case a maximum


value of the loss probability is required by HP packets, the
admission to the network has to be kept under control in order
to avoid performance degradation due to the limited resources
reserved to HP traffic.
Larger network topologies are shown in Figs. 7 and 8, where
each vertex in each graph represents an OPS node, while each
edge represents a pair of fiber links, transmitting packets in
opposite direction. Due to this assumption, the number of links
in the network is always twice the number of edges in the
corresponding graph. The two topologies present a different
connectivity, which can be quantified using the average node
degree E, defined as the average number of edges connected
to a node.
Evaluations on the above topologies have been performed
under the following assumptions:

the traffic matrix is uniform, i.e. the input traffic at the


ingress of the network is assumed to be the same on each
node and equally distributed towards all the other nodes;

Fig. 8.

Network topology with average node degree E = 5.75

this choice leads to different values of the average load


per link, according to the network topology;
the load distribution on the wavelength is balanced, i.e.
the number of wavelengths on every single link is chosen
in order to obtain the same load on each wavelength (i.e.
80%);
the amount of input traffic is the same for both the
topologies;
25% of wavelengths on each link are reserved to HP
packets;
20% of the traffic is HP;
LP traffic is re-routed with MPR strategy.
Link loss rate is shown in Figs. 9 and 10. A good level
of service differentiation has been obtained, although different
links experience a heavily unfair behavior. It is worth to notice
that the higher the topology connectivity, the higher the packet
loss rate. This is a consequence of the redistribution of the
load among a higher number of links, leading to a smaller
number of wavelengths per link and less effective wavelength
multiplexing.
As a measure of the impact of the routing strategy on
the network latency, the distribution of the number of hops
experienced by a packet is shown next. Figure 11 clearly
shows that for the less connected topology the use of the
MPR strategy involves typically alternative paths of the same
length as the shortest one. This is related to the structure of
the topology itself (similar to a Manhattan-Street Network),
which often includes several shortest paths between pairs of
nodes.
The situation is quite different for the other topology, that is
much more connected. In this case the diagonal links between

1.E00

LP loss
HP loss

1.E-01

Loss rate

1.E-02

1.E-03

1.E-04

1.E-05

1.E-06

10

15

20

25

30

35

40

45

50

Link

Fig. 9. Link loss rate for the topology of Fig. 7 with uniform traffic matrix
and balanced load distribution

1.E00

LP loss
HP loss

1.E-01

Loss rate

1.E-02

1.E-03

1.E-04

1.E-05

1.E-06

10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90
Link

Fig. 10. Link loss rate for the topology of Fig. 8 with uniform traffic matrix
and balanced load distribution

1.E+00

E = 3.25

1.E-01

E = 5.75

1.E-02

Distribution

1.E-03
1.E-04
1.E-05
1.E-06
1.E-07
1.E-08

10

11

12

13

14

15

16

17

18

19

20

No. of hops

Fig. 11. Distribution of the number of hops for the topologies of Fig. 7
(E = 3.25) and Fig. 8 (E = 5.75), with uniform traffic matrix and balanced
load distribution

nodes lead to the presence of several alternative paths crossing


one hop more than the shortest one. Furthermore, as shown
in Fig. 11, the highly congested state of several links, due to
the small number of wavelengths available, causes LP packets
to be often re-routed on alternative paths. Such packets keep
traveling inside the network, trying to follow less congested
paths before ending up to the destination node. In some
cases, they enter in temporary loops and experience a quite
large number of hops, a typical behavior of congestion-based
deflection routing techniques.
A network design procedure aiming at finding a good tradeoff between performance guarantees and network cost in terms
of number of wavelengths per link is also illustrated and
discussed in [12].
V. L INK FAILURE ISSUES
Reliability in optical networks has been a widely discussed
research topic, being a very important requirement for the next
generation optical networks [13]. Different kinds of failure can
happen even at the same time and the network must be capable
to recover from them. This must be done in an efficient way by
detecting the failure as quickly as possible, so that a recovery
procedure can be called immediately. Information loss needs
to be limited during the failure detection time, as well as when
recovery is taking place. To this purpose a smart and efficient
recovery algorithm is required.
Adaptive routing techniques could be employed in OPS
networks in the presence of optical link failures due to accidental fiber cuts or device malfunctioning. The application of
these techniques to the OPS scenario represents a completely
different approach compared to protection and restoration techniques traditionally adopted in wavelength-routed networks
[14]. The MPR strategies described above may be exploited
in order to provide reliability to the network. In fact, when
a link failure occurs, packets previously routed on that link
are transmitted on alternative paths, depending on the MPR
strategy adopted.
In this case, a key parameter is the time d required to
detect the failure and start re-routing the packets: obviously,
the shorter the fault detection delay, the smaller the number
of packets lost due to the link unavailability. Figure 12 shows
the time behavior of the packet loss related to the failed link.
Between the failure instant tf and the detection instant tf + d
the loss rate reaches a level that depends on the value of the
failure detection time d. This behavior can be easily described
by the analytical model presented in [15].
The use of MPR strategies for packet re-routing after a
link failure detection affects the traffic distribution within the
network. This is a consequence of the higher traffic delivered
to links adjacent to the one which failed. Therefore, the same
amount of overall input traffic as before is spread over a
network that is now lacking one fiber. This causes the average
load per link to be higher and, when the network works with
high loads (e.g. 0.7 Erlang/wavelength) and a failure occurs,
some links may become overloaded. In this case, a simple

Packet Loss Rate

loss rate when d = d3


loss rate when d = d2
loss rate when d = d1

0.1

loss rate without


link failure

0.01
d1

d2

failure time
0.001
0.04

0.05

tf

0.06

d3
0.07

0.08

0.09

0.1

Time

Fig. 12. Time diagram of the packet loss rate on a failed link for different
failure detection times
3.55e+12

Network Throughput [bps]

3.50e+12
3.45e+12

before failure

3.40e+12

during failure
detection

3.35e+12
3.30e+12

+73%
protection
resources

after failure
detection

no protection
resources

3.25e+12
0

0.02

0.04

0.06

Simulation Time [sec]

0.08

0.1

0.12

Fig. 13. Throughput behavior in case of failure, with and without a protection
scheme

adaptive routing approach is no longer reliable since it cannot


bring the network back to the original performance level.
This leads to the need for a specific protection scheme
based on shared resources [15], which can be realized by
dimensioning the network in two main steps. First of all,
the network is dimensioned for a given average load per
wavelength (e.g. 0.7) with relation to the input traffic matrix.
Then, for each node, further wavelengths are added to each
fiber going out of that node until it sees all its output links
with the same capacity. Simulation results for a reference case
showed that with this approach the additional cost due to
protection is 73% of the initial cost in terms of number of
wavelengths.
Figure 13 shows the trend of the throughput before and
after a failure. It can be seen that the performance level
of the failure-free scenario is almost completely recovered.
When failure occurs performance drops drastically. However,
without the protection scheme, the throughput goes even worse
after failure detection, whereas when protection is applied the
original throughput is practically restored.
VI. E FFECTS ON THE PACKET SEQUENCE
In a datagram-based communication network, packet loss as
well as out-of-order packet delivery and delay variations af-

fects end-to-end protocols behavior and may cause throughput


impairments [16], [17]. When considering TCP-based traffic
it is well known that these phenomena influence the typical
congestion control mechanisms adopted by the protocol [18]
and may result in a reduction of the transmission window
size and consequently in bandwidth under-utilization. Another
example is delay-sensitive UDP-based traffic, such as realtime traffic. In this case, because of the timing constraints,
re-transmission of lost packets is not possible and therefore
a high percentage of lost packets may result in a significant
degradation of the quality of the conversation. Moreover
unordered packets may arrive too late and/or the delay required
to reorder several out-of-sequence packets may be too high
with respect to the timing requirements of the application.
These brief and simple examples make evident the need to
limit the percentage of lost packets as well as the number
of unordered packets. In general the former event is due
to congestion while the latter is typically caused by the
fact that packets belonging to the same flow of information
can take different paths through the network and then can
experience different latency. However, in an OPS network
using the wavelength domain for congestion resolution, packets traveling along the same network path may use different
wavelengths in order to exploit wavelength multiplexing for
congestion resolution purposes, according to a given WDS
policy. Therefore it may happen that packets of the same flow
are delivered out of sequence, even though still following the
same network path.
To prove this, the first thing to do is to provide a clear
definition of out-of-sequence and a framework to evaluate the
delay jitter experienced by packets when crossing a node. A
formal definition of such a framework, that is briefly recalled
here, can be found in [19]. For a generic couple of subsequent
ordered packets Pn and Pn+1 incoming on a given OPS node,
let Jn be the jitter between them, defined as the packet offset
variation after they have crossed the node. Since the two
packets may experience different delays while crossing the
node, seven different alternatives may happen, according to
the cases shown in Fig. 14:
1) the packet sequence is always guaranteed since Pn+1
experiences more delay than Pn ;
2) the node is transparent and Pn and Pn+1 have the same
offset at the input and output (i.e. Jn = 0);
3) Pn+1 experiences less delay than Pn but at the output
it is still behind the tail of Pn ;
4) the head of Pn+1 partially overlaps the tail of Pn ;
5) Pn+1 completely overlaps Pn ;
6) Pn+1 has overtaken Pn but they are partially overlapping;
7) Pn+1 has completely overtaken Pn .
The previous formalization allows to evaluate the delay jitter
distribution as well as the amount of out-of-order packets, that
depends on the specific definition of packet sequence. For
instance, in case overlapping packets are not considered in
sequence, then the in-sequence regions will be 1, 2, and 3. If

Pn

INPUT

Pn+1
1

OUTPUT









Jitter distribution

10-1

Fig. 14.

10-2
10-3
10-4
10-5

Examples of jitter between subsequent packets

10-6
1

4
Region

Fig. 16. Delay jitter distribution for a G-type WDS algorithm with the
sequence constraint over the different regions shown in Fig. 14

10-1
10-2

1
10-1

10-3

Packet loss probability

Jitter distribution

10-2

10-4

with sequence
constraint

10-3

10-5

10-4
10-5

10-6
1

4
Region

Fig. 15. Delay jitter distribution for a G-type WDS algorithm over the
different regions shown in Fig. 14

some overlapping is allowed, then the sequence is guaranteed


also in region 4. The same for region 5, in case packets arriving
at the same time are not considered out-of-order.
The effects of a G-type WDS on packet sequence are shown
in Fig. 15, where the jitter distribution over the different
regions defined above is presented. As expected, the WDS
policy considered here causes some packets to get out of the
node unordered, since it does not take into account the correct
packet sequence when wavelength and delay are assigned to a
packet. However, the most frequent behavior is the one related
to region 2, which means that congestion happens rarely and
the packets are often transmitted transparently across the node.
In order to take into account the correct packet sequence
at the WDS stage, the G-type algorithm can be modified as
discussed in [20], introducing the additional constraint that the
current packet must not overtake the previous one. In this case
the jitter distribution is illustrated in Fig. 16. Obviously, the
trade-off behind this approach is a performance impairment,
as shown in Fig. 17 for N = 4, W = 16, B = 3.
The evaluations presented above are related only to the
packet jitter introduced by a single node due to the WDS
policy adopted. A similar approach can be applied at the whole
network level, in order to take into account also the effects on
the packet sequence of multi-path routing, whose impact may
be much stronger when the difference of the propagation times

without sequence
constraint

10-6
10-7

0.5

1.5

2.5

D (normalized to the average packet length)

Fig. 17. Impact of the sequence constraint on packet loss probability for
G-type WDS algorithm

over different paths is not negligible. The analysis of such a


scenario and the related consequences at the transport layer
are currently under investigation.
VII. C ONCLUSION
This paper presented an overview on some of the problems
arising when WDM optical packet-switched networks are
evaluated from a network-wide point of view. The effects
of resource partitioning as well as multi-path routing have
been shown by applying dynamic wavelength and buffer
management (WDS) on each link jointly with dynamic routing
strategies, both in undifferentiated and differentiated traffic
cases. The effectiveness of the proposed strategies has been
demonstrated, leading to the conclusion that the wavelength
domain is confirmed to be the key factor to achieve performance optimization. The application of adaptive routing
strategies to network recovery in case of single link failure
was also discussed, showing again the importance of the
wavelength domain to realize an effective protection scheme.
Finally, the problem of how the packet sequence may be
affected by dynamic multi-wavelength management and multipath routing was discussed.

R EFERENCES
[1] C. Qiao, M. Yoo, Optical burst switching: A new paradigm for an optical
internet, Journal of High Speed Networks, vol. 8, no. 1, pp. 69-84,
January 1999.
[2] M. J. OMahony, D. Simeonidou, D. K. Hunter, A. Tzanakaki, The application of optical packet switching in future communication networks,
IEEE Communications Magazine, vol. 39 , no. 3, pp.128-135, March
2001.
[3] W. Vanderbauwhede, D. A. Harle, Design and Modeling of an Asynchronous Optical Packet Switch for DiffServ Traffic, Proc. ONDM 2004,
Gent, Belgium, pp. 19-35, February 2004.
[4] L. Dittmann, et al., The European IST project DAVID: A viable approach
towards optical packet switching, IEEE Journal on Selected Areas in
Communications, vol. 21, no. 7, pp. 1026-1040, September 2003.
[5] F. Callegati, W. Cerroni, G. Corazza, Optimization of Wavelength
Allocation in WDM Optical Buffers, Optical Networks Magazine, vol.
2, no. 6, pp.66-72, November 2001.
[6] S. Blake, D. Black, M. Carlson, E. Davies, Z. Wang, W. Weiss, An
architecture for differentiated services, IETF RFC 2475, December
1998.
[7] D. K. Hunter, M. C. Chia, I. Andonovic, Buffering in Optical Packet
Switching, IEEE/OSA Journal of Lightwave Technology, vol. 16, no. 10,
pp. 2081-2094, December 1998.
[8] L. Tancevski et al., Optical routing of asynchronous, variable length
packets, IEEE Journal on Selected Areas in Communications, vol. 18,
no. 10, pp. 2084-2093, October 2000.
[9] F. Callegati, W. Cerroni, C. Raffaelli, P. Zaffoni, Wavelength and time
domain exploitation for QoS management in optical packet switches,
Computer Networks, vol. 44, no. 4, pp. 569-582, 15 March 2004.
[10] R. Ramamurthy, B. Mukherjee, Fixed-alternate routing and wavelength
conversion in wavelength-routed optical networks, IEEE/ACM Transactions on Networking, vol. 10, no. 3, pp. 351-367, June 2002.

[11] F. Callegati, W. Cerroni, G. Muretto, C. Raffaelli, P. Zaffoni, Adaptive


routing in DWDM optical packet switched network, Proc. ONDM 2004,
Gent, Belgium, pp. 71-86, February 2004.
[12] F. Callegati, W. Cerroni, C. Raffaelli and M. Savi, QoS differentiation
in optical packet-switched networks, to appear on Computer Communications, Elsevier, 2005.
[13] J. P. Vasseur, M. Pickavet, P. Demeester, Network Recovery: Protection and Restoration of Optical, SONET-SDH, IP and MPLS, Morgan
Kaufmann Publishers, 2004.
[14] A. Fumagalli, L. Valcarenghi, IP restoration vs. WDM protection: is
there an optimal choice?, IEEE Network, vol. 14, no. 6, pp. 34-41,
November-December 2000.
[15] F. Callegati, W. Cerroni, G. Muretto, C. Raffaelli, Exploitation of
dynamic routing for optical packet-switched network reliability, Proc.
NOC 2005, London, UK, July 2005.
[16] J. C. R. Bennett, C. Patridge, Packet reordering is not a pathological
network behavior, IEEE/ACM Transactions on Networking, vol. 7, no.
6, pp. 789-798, December 1999.
[17] M. Laor, L. Gendel, The effect of packet reordering in a backbone link
on application throughput, IEEE Network, vol. 16, no. 5, pp. 28-36,
September/October 2002.
[18] M. Allman, V. Paxson, W. Stevens, TCP congestion control IETF RFC
2581, April 1999.
[19] F. Callegati, W. Cerroni, G. Muretto, C. Raffaelli, P. Zaffoni, A framework for performance evaluation of OPS congestion resolution, Proc.
ONDM 2005, Milan, Italy, pp. 243-250, February 2005.
[20] F. Callegati, D. Careglio, W. Cerroni, G. Muretto, C. Raffaelli, J.
Sole Pareta, P. Zaffoni, Keeping the packet sequence in optical packetswitched networks, Optical Switching and Networking, vol. 2, no. 3, pp.
137-147, November 2005.