Professional Documents
Culture Documents
T
s
T
r
. (3)
From Eq. (3), we can see that the links time utilization U
t
is influenced by ring time T
r
, timeslot T
s
, and guard-
time T
g
.
BT BW
0 13
1 9
0 10
0 3
0 6
1 18
2 8
0 11
BT BW
2 8
1 9
1 18
0 13
0 10
0 3
0 6
0 11
BT BW
2 8
1 18
1 9
0 13
0 11
0 10
0 6
0 3
Random FCFS FCFS-LP
S
c
h
e
d
u
l
i
n
g
o
r
d
e
r
BT: Buffered Timeslots BW: Required Bandwidth
Fig. 6. Burst scheduling order in three burst selection
algorithms.
1048 J. OPT. COMMUN. NETW./VOL. 5, NO. 9/SEPTEMBER 2013 Zhang et al.
Furthermore, the link utilization in the frequency
domain U
b
is analyzed below. Assuming that the spectral
efficiency is 1 bpsHz, i.e., the bandwidth needed for the
transmission rate of R bps is R Hz, the average bandwidth
for one burst B is
B = LT
s
. (4)
With the guardband of B
g
, the link utilization in the
frequency domain U
b
is
U
b
=
B
B B
g
=
L
L B
g
T
s
. (5)
From Eq. (5), we can see that the links bandwidth
utilization U
b
is influenced by burst size L, timeslot T
s
,
and guardband B
g
.
Finally, the total link utilization U can be obtained from
Eqs. (3) and (5), which is stated as below:
U = U
t
U
b
=
_
T
r
T
s
T
g
_
T
s
T
r
L
L B
g
T
s
. (6)
From Eq. (6), we can see that the total link utilization U
is influenced by ring time T
r
, burst size L, timeslot T
s
,
guardtime T
g
, and guardband B
g
.
The optimal size of timeslots that achieves the maxi-
mum link utilization is analyzed. As shown in Eq. (6),
the first requirement is that T
r
(T
s
T
g
) is a positive
integer; that is,
T
s
= T
r
m T
g
; (7)
where m is an integer with 1 m < T
r
T
g
. With this
constraint, the link utilization depicted in Eq. (6) is
U =
T
r
T
s
T
g
T
s
T
r
L
L B
g
T
s
=
1
1 T
g
T
s
1
1 B
g
T
s
L
_
1
1
T
g
B
g
L
_
_
2
. (8)
The last equation holds if and only if
T
s
=
T
g
LB
g
_
. (9)
This is the second requirement to achieve the maximum
link utilization.
The analytical results for the impact of timeslot size are
shown in Fig. 8. In the analytical model, the parameters
are set as follows: T
r
= 1 ms, T
g
= 0.01 ms, L = 10 Mb,
B
g
= 10 GHz. The size of one timeslot T
s
is variable from
0.01 to 1 ms. As shown in Fig. 8, the link utilization in the
frequency domain U
b
decreases monotonically as the time-
slot size increases. The relationship between link utiliza-
tion in the time domain and timeslot size is more
complicated. Generally, the link utilization in the time do-
main increases with the timeslot size but fluctuates with
the influence of the unused time fragment that is caused
by the mismatch of round trip time T
r
and timeslot size
T
s
. The superposition of the link utilization in the time
and frequency domains results in the fact that the link uti-
lization increases with the size of the timeslot when the
timeslot is small and then decreases when the size of
the timeslot exceeds a certain level. From Fig. 8, we can
see that the optimal timeslot size with the maximum link
utilization is around 0.1 ms, which can be obtained from
Eqs. (7) and (9).
B. Number of Transmitters and Receivers
In BV-OBS ring networks, fast tunable transmitters
and receivers are needed to enable the transmission and
reception of BV bursts. The number of transmitters and
receivers are usually limited by the cost constraint. An ana-
lytical model is developed to evaluate the relationship be-
tween network throughput and the number of transmitters
and receivers.
Suppose a directional optical ring with N nodes, and
each node is equipped with M tunable transmitters and
M tunable receivers, then the total capacity of one link
is C, and the average transmission rate for one burst is
R. In each timeslot, a transmitter is only allowed to trans-
mit at most one burst and so is the receiver. Therefore, each
node can send a maximal number of M bursts to the ring in
one timeslot. The input traffic one node is allowed to send
to the ring is denoted as T. First, the input traffic T is lim-
ited by the number of transmitters:
T M R. (10)
T
r
T
s
T
s
T
s
. . .
T
u
T
g
T
g
T
g
N
r
Fig. 7. Timeslot in a BV-OBS ring network.
Fig. 8. Impact of the timeslot size in a BV-OBS ring network.
Zhang et al. VOL. 5, NO. 9/SEPTEMBER 2013/J. OPT. COMMUN. NETW. 1049
Second, the traffic load on one link should not exceed its
capacity C. With uniform traffic assumption, the traffic
T is uniformly distributed to the remaining N 1 nodes,
and the traffic load on one link is composed of
N(N 1)2 node pairs, which indicates that
T
N 1
N(N 1)
2
C. (11)
Therefore, the input traffic for each node is
T min(M R; 2CN). (12)
With the sufficient number of transmitters, the input traf-
fic for each node can achieve its maximum T
m
:
T
m
= 2CN. (13)
The sufficient number of transmitters that can fully
utilize the link capacity is
M
Tx
=
_
2C
N R
_
; (14)
where {x represents the minimal integer that is not
smaller than x.
The discussions above are focused on the input traffic
for one node. In the proposed collision-free BV-OBS ring
network, bursts will never be blocked while transmitting
in the ring. Nevertheless, the bursts may not be success-
fully received because there are not enough receivers at
the destination nodes. For each node, the total incoming
traffic is T + N2 as indicated in Eq. (11), of which T traffic
is to drop at this node. Therefore, the probability that an
incoming burst is a dropping burst is
p =
T
T N2
= 2N. (15)
As the average transmission rate for one burst is R, the
total number of incoming bursts is approximately
n =
_
N T
2R
_
; (16)
where [x] is the nearest integer of x. Among the n incoming
bursts, each burst has a probability of 2N to be dropped at
the node. With the independence assumption of burst des-
tinations, the probability that k bursts are to be dropped at
the node can be calculated by the Bernoulli formula:
P(k) = C
k
n
p
k
(1 p)
nk
. (17)
As the node can only receive a maximal number of M bursts
in one timeslot, if the number of dropping bursts k is larger
than M, the remaining k M bursts will be blocked. The
average burst blocking probability because of the limita-
tion on the number of tunable receivers is
P
r
=
R
T
n
k=M1
(k M) P(k); (18)
where
n
k=M1
(k M) P(k) is the average number of
dropping bursts blocked in the node in one timeslot. When
the number of receivers M is sufficiently large, the burst
blocking probability P
r
would tend to zero.
The network throughput is defined as the maximum
traffic successfully transmitted through the ring, which is
calculated as the total input traffic minus the blocked traffic:
= N T (1 P
r
). (19)
With a sufficient number of transmitters and receivers,
the input traffic can achieve its maximum T
m
and the
reception blocking probability at the destination nodes
approaches zero. In this case, the maximum network
throughput is
m
= N T
m
= 2C. (20)
The above analytical model is developed with the
assumption of uniform traffic distribution. In the next
section, we will verify it by simulation with both uniform
and nonuniform traffic.
VI. PERFORMANCE EVALUATION
In this section, the performance of the proposed collision-
free BV-OBS ring network is evaluated by simulation,
including network throughput, delay, buffer size, and
service fairness.
A. Number of Transmitters and Receivers
As revealed in the previous section, the network
throughput of a BV-OBS ring network is related to the node
number and the number of transmitters and receivers for
each node. In this subsection, we verify the analytical re-
sults by simulation. In the simulation, the link capacity C
is 4 Tbps and the average transmission rate for each burst
R is 100 Gbps. We assume that there are infinite bursts at
the local buffers waiting to be transmitted to the ring net-
work. Therefore, the network capacity is limited by the link
capacity and the number of transmitters and receivers.
The network throughput with the number of transmit-
ters and receivers is shown in Fig. 9. Three ring networks
with node number N = 6, 10, and 15 are compared. The
traffic load is uniformly distributed among all node pairs.
From Fig. 9, we can see that the analytical results (filled
symbols) fit with the simulation results (unfilled symbols)
quite well. The simulation results are a bit lower than
the analytical results when the network throughput ap-
proaches its maximum since the link capacity cannot be
fully utilized because of spectral fragmentation. When the
number of transmitters and receivers is small, the network
throughput increases with the node number N. The reason
is that the input traffic at each node is mainly limited by
1050 J. OPT. COMMUN. NETW./VOL. 5, NO. 9/SEPTEMBER 2013 Zhang et al.
the number of transmitters and receivers when it is small.
When the number of transmitters and receivers is suffi-
ciently large, the network throughput is mainly limited
by the link capacity and is almost the same for different
node numbers N as shown in Eq. (20).
With a given number of transmitters and receivers, the
network throughput is also affected by traffic distribution.
As we assume there is always enough traffic at the source
nodes, the traffic amount that can be successfully received
at the destination nodes contributes to the network
throughput and also decides the amount of traffic that
can be added to the ring. Therefore, we simulate the
scenario with nonuniformtraffic to different destination no-
des. We assume that the traffic loads destined for odd nodes
are larger than even nodes. The ratio of the two is defined
as the traffic uniform ratio (TUR). The traffic is uniformly
distributed when the TUR is 1. With a larger value of the
TUR, the traffic is more nonuniformly distributed.
The network throughput with nonuniform traffic
distribution is shown in Fig. 10. In the simulation, the node
number N is fixed to 10 and the TUR varies from 1 to 5.
From Fig. 10, we can see that with the same number
of transmitters and receivers, the network throughput
decreases when the TUR increases; i.e., the traffic is more
nonuniformly distributed. This is because, when traffic is
nonuniformly distributed, more receivers are required
for nodes with more traffic loads.
The sufficient number of transmitters and receivers is
defined as the value at which the network throughput
achieves 90% of its maximum as shown in Eq. (20). The
simulation results of a sufficient number of transmitters
and receivers with different node numbers N and TUR
are shown in Table II. With uniform traffic (TUR = 1),
the sufficient number of transmitters and receivers de-
creases when the node number increases, which is consis-
tent with Eq. (14). The sufficient number of transmitters
and receivers with nonuniform traffic is larger than that
with uniform traffic and increases with the value of
the TUR.
B. Performance Comparison of Collision-Free OBS
Rings
In this subsection, the network throughput and end-
to-end delay of five collision-free OBS ring networks are
compared: UPS, SBCT, LightRing, CORNet, and the
proposed BV-OBS. The first four are proposed in a WDM
network, and the last one is proposed in an EON. For a fair
comparison, the link capacity in the WDM network and in
the EON is the same, with a total rate of 4 Tbps. The link
consists of 40 wavelengths at the rate of 100 Gbps in the
WDM network and 400 subbands at the rate of 10 Gbps
in the EON. Since a small or even zero guardband is re-
quired for each subband in the EON with orthogonal band
multiplexed OFDM (OBM-OFDM) [26], the influence of
the guardband is not considered, and it is set to zero in
the simulation. We assume there is a sufficient number
(M = 20) of transmitters and receivers for each node.
The basic simulation scenario is set as follows. The net-
work topology is a directional ring network with six nodes
and its round trip time is 1 ms, i.e., the circumference of
the ring is 200 km. Uniform traffic distribution and burst
size are assumed. The burst size varies from 0 to 20 Mb
with an average of 10 Mb. In BV-OBS ring networks, the
timeslot is set to 0.1 ms, thus the bandwidth for one
burst varies from 1 to 20 subbands. In WDM-based ring
networks, the bandwidth for one burst is fixed to one wave-
length (at the rate of 100 Gbps), and the burst duration
varies from 0 to 0.2 ms.
Fig. 9. Network throughput versus number of transmitters and
receivers with uniform traffic.
Fig. 10. Network throughput versus number of transmitters and
receivers with nonuniform traffic.
TABLE II
SUFFICIENT NUMBER OF TRANSMITTERS AND RECEIVERS
WITH DIFFERENT NODE NUMBER N AND TUR
TUR
N 1 2 3 4 5
6 14 17 19 20 21
10 11 12 14 15 15
15 9 11 12 13 13
Zhang et al. VOL. 5, NO. 9/SEPTEMBER 2013/J. OPT. COMMUN. NETW. 1051
The network throughputs of the five OBS rings in the
basic scenario are shown in Fig. 11. The throughput is
normalized to the maximum network throughput defined
in Eq. (20). The input traffic load is also normalized to
the maximum load in Eq. (13). As we can see, the through-
puts of SBCT and BV-OBS are the largest, which can
achieve 1 when the input traffic load is larger than 1. How-
ever, the throughputs of the other three protocols are all
less than 1. The maximum throughputs of UPS, LightRing,
and CORNet are 0.51, 0.1, and 0.5 accordingly. The
throughput of UPS is limited because of the wasted band-
width occupied by incomplete burst segments. When the
input traffic load is larger than 0.51, the throughput drops
to almost 0 because the network is in unstable self-exciting
status, where more traffic will result in higher probability
of interrupted transmission and higher probability of
interrupted transmission will induce even more traffic
by burst retransmission. As a result, the traffic in the ring
network would be full of incomplete burst segments and no
burst would be transmitted successfully. The throughput of
LightRing is limited by the ratio of the average burst du-
ration to the round trip time. The throughput is very low in
this simulation scenario because the average burst dura-
tion is only 10% of the round trip time. The LightRing pro-
tocol is more bandwidth efficient when the average burst
duration is close to or larger than the rings round trip time.
The throughput of CORNet is limited by the maximum to
average ratio (MAR) of the burst size, and it can be im-
proved if the distribution of burst sizes is more concen-
trated to the average value.
As mentioned above, the network throughputs of
LightRing and CORNet are highly dependent on the burst
size distribution. The network throughputs of LightRing,
CORNet, and BV-OBS rings with different burst size
distributions are shown in Fig. 12. The average burst size
in LightRing is set to 100 Mb so that the average burst
duration is 1 ms, which is the same as the rings round trip
time. The average burst size in CORNet and BV-OBS
rings still remains 10 Mb. The parameter of MAR is used
to represent the different burst size distributions. In the
simulation, the MAR varies from 1 to 5 at the step of 0.5.
The burst size distribution is more dispersive with a larger
MAR. As shown in Fig. 12, the network throughputs of
LightRing, CORNet, and BV-OBS achieve their maximum
when the MARis 1 (i.e., bursts are a fixed size). The network
throughputs of LightRing and CORNet decrease rapidly
when the value of the MAR increases. On the contrary,
the network throughput of a BV-OBS ring can always
achieve its maximum whatever the value of the MAR is.
This is because the bandwidth occupation time in LightRing
and CORNet cannot be flexibly allocated in the time do-
main, while the bandwidth for bursts in a BV-OBS ring
can be flexibly allocated in the frequency domain. With
the same value of the MAR, the network throughput of
LightRing is higher than that of CORNet. The reason is
that the bandwidth occupation time for one burst is fixed
in CORNet but can be variable with integer multiples of
the round trip time in the LightRing protocol.
The average end-to-end delays of the five OBS rings in
the basic scenario are shown in Fig. 13. The end-to-end
delay is composed of the propagation delay, offset time,
Fig. 11. Network throughput of collision-free OBS ring networks.
Fig. 12. Network throughput of LightRing, CORNet, and BV-
OBS ring networks with different burst size distributions.
Fig. 13. End-to-end delay of collision-free OBS ring networks.
1052 J. OPT. COMMUN. NETW./VOL. 5, NO. 9/SEPTEMBER 2013 Zhang et al.
and queuing delay in the local buffer, while the assembling
time is not considered. The average propagation delay in
the ring network is 0.5 ms, i.e., half of the round trip time
in the ring. As shown in Fig. 13, the end-to-end delay of
SBCT is the largest in the five protocols, which is mainly
caused by the extra offset time introduced to provide higher
priority to bypass bursts. The end-to-end delays of the
other four protocols are all closed to the propagation delay.
The queuing delay is negligible when the traffic load is low
and increases rapidly when the load approaches its limit.
For example, the queuing delay of a BV-OBS ring network
is no more than 10%of the propagation delay when the nor-
malized traffic load is lower than 0.9.
Moreover, the delay of the SBCT protocol is highly de-
pendent on the node number N and the maximum burst
duration. The end-to-end delays of SBCT and BV-OBS
rings with different burst size distributions and node num-
bers are shown in Fig. 14. In the simulation, the average
burst size is still set to 10 Mb, while the maximum burst
size varies from 10 to 100 Mb. In a BV-OBS ring network,
the burst duration is fixed to 0.1 ms and the burst band-
width can be variable according to the burst size. However,
in the SBCT protocol the burst bandwidth is fixed and the
burst duration is variable according to the burst size. In
other words, the maximum burst duration in the SBCT
protocol should vary from 0.1 to 1 ms to support the differ-
ent burst size distributions. The normalized traffic load is
set to 0.8 so that the queuing delay is neglectable. As shown
in Fig. 14, the delay of the SBCT protocol increases dra-
matically with the maximum burst size and node number
N, while the delay of the BV-OBS ring always remains low.
For example, the delay of the SBCT protocol is almost 8 ms
with the maximum burst size equal to 100 Mb and node
number N = 15, while that of the BV-OBS ring is only
0.5 ms (i.e., the basic propagation delay). The delay of
SBCT is nearly 15 times larger than that of the BV-OBS
ring. The delay of SBCT could be even higher with more
node numbers in the ring, which implies that the applica-
tion of the SBCT protocol is limited to small-scale rings.
Furthermore, the extra delay in SBCT increases with
the burst hop; i.e., bursts with larger hops would suffer
from even longer extra delays. It will deteriorate the
service unfairness of bursts with different hops and may
not be tolerable for some delay-sensitive or delay-jitter-
sensitive traffic.
In summary, we verify that the proposed BV-OBS ring
network outperforms the previous WDM-based OBS ring
networks by achieving both high network throughput and
low end-to-end delay. As bandwidth-variable transmitters/
receivers (BV-Tx/Rx) and BV-ROADM are required in
BV-OBS rings, the cost may be higher than the WDM-
based OBS rings. However, it is the trend that optical net-
works should be more flexible and the cost of BV devices
would be reduced in the future. Furthermore, with smaller
guardband and higher order modulation formats, the
spectrum efficiency in EONs can be much higher than that
in WDM networks. The high spectrum efficiency of
14 bitssHz with 256 QAM-OFDM has been demon-
strated in [27]. Therefore, we believe that BV-OBS ring
networks would be more cost-effective than the previous
WDM-based OBS ring networks, especially when the spec-
trum efficiency is considered.
C. Performances of Waveband Assignment
Algorithms
In this subsection, we investigate the performances of
the proposed MA waveband assignment algorithm, includ-
ing average queuing delay and average buffer size, and
compare it with the random and FF algorithms.
The average queuing delays of the three waveband as-
signment algorithms are shown in Fig. 15. As the queuing
delay for light and median traffic load is negligible (less
than 10
3
ms), we only plot the data with traffic load larger
than 0.8. From Fig. 15, we can see that the delay improve-
ment in the MA algorithm is over 20% compared with the
FF algorithm and over 50% compared with the random
algorithm. For example, the delays of the MA, FF, and
Fig. 14. End-to-end delay of SBCT and BV-OBS ring networks
with different burst size distributions and node numbers.
Fig. 15. Average queuing delay of waveband assignment
algorithms.
Zhang et al. VOL. 5, NO. 9/SEPTEMBER 2013/J. OPT. COMMUN. NETW. 1053
random algorithms at load = 0.96 are 0.288, 0.441, and
0.587 ms, respectively, with reduction of approximately
25% and 51%. Moreover, with = 0.9, the delays of the
three algorithms are 0.0507, 0.0806, and 0.115 ms, and
nearly 30% and 56% improvement are achieved.
The average buffer sizes of the three waveband assign-
ment algorithms are shown in Fig. 16. As the volume of the
local buffer is assumed to be infinite, adding bursts that
cannot find available wavebands at one timeslot would
be buffered and try again in the following timeslots. The
average buffer size is defined as the average amount of
bursts waiting in the local electrical buffer in one node after
the scheduling of adding bursts. The trend of buffer size in
terms of traffic load is similar to that of queuing delay
shown in Fig. 15. The buffer size is the smallest in the MA
algorithm and the largest in the random algorithm, while
the FF is in the middle. We notice that the improvement
by the MA algorithm in buffer size is more significant than
that in queuing delay. For example, with load = 0.96, the
buffer size reduces from 145 MB in the random algorithm
to 65.5 MB in the MA algorithm, with improvement of
almost 57%, which is larger than the 51% reduction in
queuing delay. It is reasonable because the proportion of
bursts with larger size in the buffer is larger than that
with small size as the former have more difficulty finding
available wavebands.
D. Fairness of Bursts With Different Bandwidths
The fairness of bursts with different bandwidths is
evaluated in this subsection. First, a fairness coefficient
is defined to evaluate the fairness of bursts with different
bandwidths. As the proposed BV-OBS ring is collision free,
the queuing delay is chosen as the performance metric. In
the simulation, the bandwidth of one burst varies from 1 to
20 subbands. The average queuing delays of bursts with
these 20 different bandwidths are collected. The standard
deviation and average value of these 20 queuing delays
are calculated, and the fairness coefficient is defined as
the ratio of standard deviation to average value. Generally,
the fairness coefficient is a positive value, and the smaller
it is, the more fairness is achieved. It is called complete
fairness if the fairness coefficient is 0, i.e., the average
queuing delays of bursts with the 20 different bandwidths
are all the same.
The fairness coefficients of the three burst selection
algorithms are shown in Fig. 17. The fairness coefficient
of the random algorithm is the largest (about 0.8), while
that of the FCFS-LP algorithm is the smallest (about
0.45). The fairness coefficient of the FCFS algorithm is
in the middle and is dependent on the traffic load. The fair-
ness coefficient of the FCFS algorithm is close to that of the
FCFS-LP algorithm at high traffic load ( = 0.96) and is
close to the randomalgorithmwhen the traffic load is lower
( = 0.8). In the FCFS algorithm, bursts are selected in or-
der of the number of timeslots they have been in the buffer.
If all the bursts in the buffer have the same buffered time-
slots, the FCFS algorithm will be equivalent to the random
algorithm. Therefore, the FCFS algorithm is close to the
random algorithm at lower traffic load as few bursts are
needed to be buffered; i.e., the buffered timeslots for most
bursts are zero. The FCFS-LP algorithm is better than the
FCFS algorithm in that bursts with the same buffered
0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96
0
30
60
90
120
150
Normalized Load
A
v
e
r
a
g
e
B
u
f
f
e
r
S
i
z
e
/
M
B
Random
FF
MA
Fig. 16. Average buffer size of waveband assignment algorithms.
Fig. 17. Fairness coefficient of buffer selection algorithms.
2 4 6 8 10 12 14 16 18 20
0
0.02
0.04
0.06
0.08
0.1
0.12
Burst Bandwidth /subbands
A
v
e
r
a
g
e
Q
u
e
u
i
n
g
D
e
l
a
y
/
m
s
Random
FCFS
FCFS-LP
Fig. 18. Average queuing delay versus burst bandwidth at traffic
load = 0.9.
1054 J. OPT. COMMUN. NETW./VOL. 5, NO. 9/SEPTEMBER 2013 Zhang et al.
timeslots are further sorted by their bandwidths and those
with larger bandwidths are first selected.
The average queuing delays of the three burst selection
algorithms at load = 0.9 are shown in Fig. 18. We can
see that the delay increases with the burst bandwidth
monotonously in both the random and FCFS algorithms.
This is the unfairness of bursts with different bandwidths,
as neither of the two algorithms considers the different
bandwidths of bursts. Among the two algorithms, the delay
disparity of bursts in the FCFS algorithm is smaller than
that in the random algorithm, as the waiting time of one
burst in the buffer is taken into consideration in the FCFS
algorithm. The fairness is further improved in the FCFS-
LP algorithm as both the waiting time and the burst band-
width are considered. As shown in Fig. 18, the delay of
bursts with large bandwidth is reduced and that with small
bandwidth is increased. As a result, the delays of bursts
with different bandwidths are more even in the FCFS-
LP algorithm, although the delay of bursts with medium
bandwidths is a bit higher.
VII. CONCLUSION
In this paper, for the first time, we introduced flexible
bandwidth allocation to an OBS ring network and proposed
novel BV-OBS to construct a collision-free, low-latency, and
bandwidth-efficient optical ring network. The spectral
fragmentation and unfairness of bursts with different
bandwidths are two problems newly emerged in BV-OBS
ring networks. The MA waveband assignment algorithm
is proposed to minimize the spectral fragmentation, and
the FCFS-LP burst selection algorithm is proposed to
improve the fairness.
Simulation results show that the proposed BV-OBS ring
network can achieve the maximum network throughput
while the end-to-end latency remains low. The increment
of end-to-end delay is less than 0.05 ms (or approximately
10% of the average propagation delay) when the normal-
ized traffic load is lower than 0.9. Among the three wave-
band assignment algorithms, random, FF, and MA,
simulation results show that the MA algorithm has the
lowest queuing delay and the smallest buffer size, with im-
provement of more than 50% over the random algorithm
and more than 20% over the FFalgorithm. For burst selec-
tion algorithms, the FCFS-LP algorithm achieves the best
fairness of bursts with variable bandwidths, in which the
queuing delays of bursts with small or large bandwidths
are low while those with medium bandwidths are a little
higher.
ACKNOWLEDGMENT
This work was supported by the National Basic Research
Program of China (973 Program, Nos. 2010CB328201 and
2010CB328202), the National Natural Science Foundation
of China (NSFC, Nos. 60907030, 60736003, 61205058,
60931160439), the National Hi-tech Research and
Development Program of China (No. 2011AA01A106),
and the Open Fund of the State Key Laboratory of
Information Photonics and Optical Communications
(Beijing University of Posts and Telecommunications),
China.
REFERENCES
[1] C. M. Qiao and M. S. Yoo, Optical burst switching (OBS)A
new paradigm for an optical Internet, J. High Speed Netw.,
vol. 8, pp. 6984, 1999.
[2] M. Yoo and C. Qiao, Just-enough-time (JET): A high speed
protocol for bursty traffic in optical networks, in IEEE/
LEOS Technologies for a Global Information Infrastructure,
1997, pp. 2627.
[3] J. Y. Wei and R. I. McFarland, Just-in-time signaling for
WDM optical burst switching networks, J. Lightwave
Technol., vol. 18, pp. 20192037, Dec. 2000.
[4] L.-M. Peng and Y.-C. Kim, Investigation of the design of
MACprotocols for TT-TR-based WDMburst-switched ring net-
works, J. Opt. Commun. Netw., vol. 1, pp. A25A34, 2009.
[5] N. Deng, Q. Xue, M. Li, A. Lord, P. Willis, S. Cao, and Z. Feng,
Network modelling and techno-economic analysis of optical
burst ring for metropolitan applications, in Proc. European
Conf. and Exhibition on Optical Communication (ECOC),
Amsterdam, The Netherlands, 2012, paper P5.07.
[6] C. Li, N. Deng, M. Li, Q. Xue, and P. Wai, Performance
analysis and experimental demonstration of a novel network
architecture using optical burst rings for interpod communi-
cations in data centers, IEEE J. Sel. Top. QuantumElectron.,
vol. 19, no. 2, 3700508, 2012.
[7] L. M. Peng, C. H. Youn, W. Tang, and C. M. Qiao, A novel
approach to optical switching for intradatacenter network-
ing, J. Lightwave Technol., vol. 30, no. 2, pp. 252266,
Jan. 2012.
[8] N. Deng, Q. Xue, M. Li, G. Gong, and C. Qiao, An optical
multi-ring burst network for a data center, in Optical
Fiber Communication Conf. (OFC), Anaheim, CA, 2013,
paper OTh1A.5.
[9] L. Xu, H. G. Perros, and G. N. Rouskas, A simulation study of
optical burst switching and access protocols for WDM ring
networks, Comput. Netw., vol. 41, no. 2, pp. 143160,
Feb. 2003.
[10] Y. Arakawa, N. Yamanaka, and I. Sasase, Optical burst
switched ring network with upstream prioritized switching
and distributed fairness control, Electron. Commun. Jpn.,
vol. 90, pp. 3039, 2007.
[11] J. P. Park and M. S. Lee, Simultaneous burst and burst
control packet transmission protocol for optical burst
switching ring networks, ETRI J., vol. 29, pp. 116119,
2007.
[12] A. Fumagalli and P. Krishnamoorthy, A low-latency and
bandwidth-efficient distributed optical burst switching
architecture for metro ring, in Proc. ICC, Alaska, May 2003,
vol. 2, pp. 13401344.
[13] H.-T. Lin and W.-R. Chang, CORNet: A scalable and band-
width-efficient optical burst switching ring architecture for
metro area networks, in Proc. of 2006 Int. Conf. on Network-
ing and Services, California, July 2006.
[14] J. Triay and C. Cervello-Pastor, Distributed contention
avoidance in optical burst-switched ring networks, in Proc.
of 11th IEEE Int. Conf. on Communication Systems (ICCS),
Singapore, 2008, pp. 715720.
[15] A. Ge, F. Callegati, and L. S. Tamil, On optical burst switch-
ing and self-similar traffic, IEEE Commun. Lett., vol. 4,
pp. 98100, Mar. 2000.
Zhang et al. VOL. 5, NO. 9/SEPTEMBER 2013/J. OPT. COMMUN. NETW. 1055
[16] V. M. Vokkarane, K. Haridoss, and J. P. Jue, Threshold-based
burst assembly policies for QoS support in optical burst-
switched networks, in Proc. SPIE, vol. 4874, pp. 125136,
July 2002.
[17] S. Oh, H. H. Hong, and M. Kang, A data burst assembly
algorithm in optical burst switching networks, ETRI J.,
vol. 24, no. 4, pp. 311322, 2002.
[18] J. Armstrong, OFDM for optical communications, J. Light-
wave Technol., vol. 27, pp. 189204, 2009.
[19] C. Zhao, Y. Chen, S. Zhang, J. Li, F. Zhang, L. Zhu, and Z.
Chen, Experimental demonstration of 1.08 Tbs PDM
CO-SCFDMtransmission over 3170 kmSSMF, Opt. Express,
vol. 20, no. 2, pp. 787793, 2012.
[20] M. Jinno, H. Takara, B. Kozicki, Y. Tsukishima, Y. Sone, and
S. Matsuoka, Spectrum-efficient and scalable elastic
optical path network: Architecture, benefits, and enabling
technologies, IEEE Commun. Mag., vol. 47, pp. 6673, 2009.
[21] S. Thiagarajan, M. Frankel, and D. Boertjes, Spectrum effi-
cient super-channels in dynamic flexible grid networksA
blocking analysis, in Optical Fiber Communication Conf.
(OFC), Los Angeles, CA, 2011, paper OTuI6.
[22] A. Castro, L. Velasco, M. Ruiz, M. A. Klinkowski, J. P.
Fernndez-Palacios, and D. Careglio, Dynamic routing and
spectrum (re)allocation in future flexgrid optical networks,
Comput. Netw., vol. 56, pp. 28692883, 2012.
[23] Y. Yin, K. Wen, D. J. Geisler, R. Liu, and S. Yoo, Dynamic
on-demand defragmentation in flexible bandwidth elastic
optical networks, Opt. Express, vol. 20, pp. 17981804,
2012.
[24] R. Dischler, F. Buchali, and A. Klekamp, Demonstration
of bit rate variable ROADMfunctionality on an optical OFDM
superchannel, in Optical Fiber Communication Conf. (OFC),
San Diego, CA, Mar. 2010, paper OTuM7.
[25] Y. Chen, J. Li, C. Zhao, L. Zhu, F. Zhang, Y. He, and Z. Chen,
Experimental demonstration of ROADM functionality on an
optical SCFDM superchannel, IEEE Photon. Technol. Lett.,
vol. 24, pp. 215217, 2012.
[26] Y. Qi, T. Yan, M. Yiran, and W. Shieh, Experimental demon-
stration and numerical simulation of 107 Gbs high spectral
efficiency coherent optical OFDM, J. Lightwave Technol.,
vol. 27, pp. 168176, 2009.
[27] T. Omiya, M. Yoshida, and M. Nakazawa, 400 Gbits 256
QAM-OFDM transmission over 720 km with a 14 bitsHz
spectral efficiency by using high-resolution FDE, Opt.
Express, vol. 21, no. 3, pp. 26322641, 2013.
1056 J. OPT. COMMUN. NETW./VOL. 5, NO. 9/SEPTEMBER 2013 Zhang et al.