IJCSNS International Journal of Computer Science and Network Security, VOL.6 No.

8A, August 2006 197

SPACES: A Distributed Fairness Algorithm for Bus-Based

Metropolitan Optical Network
Daniel POPA and Tülin ATMACA,

Institut National des Télécommunications, Dept. RST

9 rue Charles Fourier, 91011 Evry, France
Summary protocols maintain synchronous slot boundaries over many
In metropolitan optical networks, the fairness issue has mostly wavelengths through dispersion compensation. Next, an
been analyzed with regard to synchronous transmission (i.e., access protocol based on asynchronous CSMA/CA
slotted WDM ring). In this context, we focus on the performance technique shares efficiently the available capacity between
of asynchronous CSMA/CA protocol with variable packet sizes. access points at the time scale of individual packets.
We examine the fairness issue arising between ring nodes
Finally, when the asynchronous CSMA/CA protocol is
sharing a common data channel and show that the asynchronous
CSMA/CA protocol suffers from bandwidth fragmentation and used in passive optical ring configuration costs decrease
positional priority fairness problems. The paper proves that the significantly (e.g., [10]).
phenomenon of bandwidth fragmentation represents the most
important issue in asynchronous CSMA/CA protocol because it In this article, we focus on the performance issue in
leads to bandwidth waste and, thus, increases unfairness for DBORN. The DBORN architecture will be briefly
certain ring nodes. In this regard, we explore limitations of a described in this paper, but for more details the reader is
previous proposed solution, the so-called Traffic Control invited to refer to [7]. The performance of asynchronous
Architecture using Remote Descriptors (TCARD). Afterwards, optical CSMA/CA technique relies mainly on how optical
we present an algorithm, called Smart sPACEd tranSmission
resource (i.e., channel/wavelength) is shared among
(SPACES), which deals with the aforementioned fairness
problems. The devised scheme is based on a statistical and fully different access nodes. Even though the MAC protocol
distributed mechanism to grant fair access to shared resources. enables collision-free transmission over the shared
As illustrated in this article, SPACES algorithm alleviates medium, it does not address the inherent fairness issue
performance degradation and resources underutilization while which is pronounced in the case of shared medium ring
achieving fairness among ring nodes. networks. Previous analyses on performance of shared
medium networks with ring topology have polarized the
Key words: research community to the major problem of such
Asynchronous optical CSMA/CA protocol, Variable length configuration: The performance of a node depends upon
packet, Shared bus ring network, Distributed fairness algorithm, its physical location within the network. The relative
Bandwidth fragmentation, Performance.
performance between ring nodes was shown to be quite
unfair.
1. Introduction
Unfortunately, however, the fairness issue is more
A crucial aspect of networking in metropolitan area is the complex in ring configuration with CSMA/CA protocol.
design of a robust and simple access scheme for building The fairness-related problems, as we shall show in this
bandwidth-efficient and flexible metropolitan optical article, are positional priority and bandwidth
networks [11][12]. The asynchronous CSMA/CA scheme fragmentation respectively. Positional priority is a
was proposed in two recent experimental projects: Dual- fairness issue common to all ring architectures (i.e., slotted
Bus Optical Ring Network (DBORN) [7][10] and Hybrid or unslotted, with or without spatial reuse of the
Optical Ring NETwork (HORNET) [5], respectively. As it bandwidth). Because the priority is given to upstream
appears, the asynchronous CSMA/CA scheme is a good traffic in this access scheme, a problem known as
candidate for access protocol in next generation of “starvation” [14] can happen when downstream nodes are
metropolitan optical networks. Several arguments come in constantly covered by upstream traffic and not able to
favor of this affirmation. First, it can be a fully-distributed access network for a very long period of time. In
asynchronous protocol not based on a centralized CSMA/CA protocol, positional priority is more important
controller or separate control wavelength to synchronize because the spatial reuse of the bandwidth is not possible.
the operation of nodes on the ring. This is an advantage in Bandwidth fragmentation appears as a direct consequence
implementation compared to slotted optical MAC of the random and asynchronous insertion of optical
protocols. In WDM ring networks (e.g., [3]), slotted MAC frames among ring nodes: Upstream nodes create useless

Manuscript received August 5, 2006.

Manuscript revised August 25, 2006.
198 IJCSNS International Journal of Computer Science and Network Security, VOL.6 No.8A, August 2006

idle periods that increase fairness problems at downstream to deal with the fairness issue in DBORN. Our solution,
nodes. There are only two papers available identifying the called Smart sPACEd tranSmission (SPACES), is based
phenomenon of bandwidth fragmentation as a factor of on a “smart” smoothing mechanism that aims dealing with
unfairness in asynchronous CSMA/CA protocol. We both bandwidth fragmentation and positional priority
highlight, however, that previous analytical works [1][2] fairness-related problems. SPACES regulates the
do not provide any detail how this phenomenon influences asynchronous transmission of optical frames by using two
the performance of the access protocol. The complexity of timers to control the process of emission. The performance
the developed mathematical models shows that this issue evaluation study shows that this algorithm deals efficiently
is complicated by the lack of possibility to analytically with aforementioned fairness problems. As illustrated in
describe such process and, thus, of sufficiently accurate this paper, the proposed solution achieves performance
knowledge of distribution of voids among ring nodes. and fairness among ring nodes.
Therefore, we use simulation models to provide an
intuitive explanation of the interaction between bandwidth The remaining parts of this article are organized as follows.
fragmentation and the fairness issue in DBORN. In Section 2, we present the basic principles of the
Furthermore, we prove that bandwidth fragmentation network architecture. The fairness issue is discussed in
represents the most important issue in such architecture. Section 3. Afterwards, we explain, in Section 4, why the
traditional fairness algorithms cannot perform
Most of the schemes proposed to deal with the starvation satisfactorily in ring topology with CSMA/CA scheme.
problem in ring configurations use global fairness Additionally, we explore limitations of the TCARD
algorithms. Existing mechanisms proposed in MAGNET mechanism. The principles of the SPACES scheme are
[17], Orwell [18] and ATMR [15] are all global fairness explained in Section 5. In Section 6, we show through
algorithms. Such algorithms operate using fairness cycles simulation results that the SPACES algorithm achieves
and allow nodes to enter the new cycle after an idle period fairness and improves the performance of the access
of time between successive fairness cycles. The local protocol. Finally, conclusions are provided in Section 7.
fairness algorithms put a quota on the transmission of
upstream nodes in order to preserve enough bandwidth for
downstream nodes. Some of proposed schemes (e.g., 2. The Network Architecture
MetaRing [14]) regulate access through a token-based-like
approach, since others (i.e., Token Bucket) simply survey
that access nodes do not violate the negotiated throughput. The network architecture is depicted in Fig. 1. The system
However, as we shall explain latter in this article, is organized around a hub which connects the metro
aforementioned fairness control mechanisms do not network to the backbone. Two counter rotating optical
perform satisfactorily in ring configuration with fibers (for protection reasons) with transparent and passive
CSMA/CA scheme. Optical Add Drop Multiplexers (OADMs) compose the
network transmission medium.
Recently, Ciavaglia et al. [2][6] proposed the so-called
Traffic Control Architecture using Remote Descriptors The architecture is based on two buses: an upstream and a
(TCARD) to alleviate the fairness issue in DBORN. downstream bus, respectively. The upstream bus is
Although this scheme deals with the starvation problem at composed of one or several wavelength(s) created in edge
light-to-moderate load, we present in this paper nodes and terminated at the hub node. Downstream
unexpected behavior of TCARD algorithm. For moderate- wavelengths start from transmitters in the hub and are
to-high load, starvation moves backward to upstream terminated in ring nodes. Several edge nodes share
nodes. upstream and downstream channels respectively in
asynchronous time division multiplexing. An access node
The design of a scalable fairness algorithm is not only can be attached to more than one upstream or downstream
limited to the performance of the access protocol, but also wavelength. A simple collision avoidance mechanism is
to traffic dynamicity. The parameters setting of TCARD in implemented through photodiode power detection on each
any case relies on the information regarding the average of locally accessible upstream wavelength. Thanks to a
the offered traffic load in downstream nodes. In real sliding window realized by a photodiode and a Fiber
network operation, due to traffic variation, is not always Delay Line (FDL), ring nodes detect idle periods and
easy to predict precisely the load of the best effort traffic capitalize adequate null periods. A size of FDL slightly
in edge nodes and, thus, to collect information for setting larger than the network Maximum Transmission Unit
of the TCARD parameters. In this context, we introduce a (MTU) ensures packet insertion on upstream wavelengths
statistical and fully-distributed traffic control mechanism without collision. When the null period is too small, the
IJCSNS International Journal of Computer Science and Network Security, VOL.6 No.8A, August 2006 199

information remains electronically buffered in ring nodes depends on the activity of the first node. Consequently, the
until a sufficient void space is detected. process of transmission at jth node (j = 2,..., N) depends on
the activity of previous j-1 nodes. When upstream nodes
constantly occupy the transmission media, starvation takes
place at downstream nodes. In addition, access protocol
cannot provide concurrent transmission over distinct
segments of the ring (i.e., spatial reuse of the bandwidth),
because the hub is the only one which remove frames from
the upstream channels. That is, in such architecture,
starvation represents a fairness issue more important than
in other ring configurations.

Fig.1: The network architecture

It is obvious that this architecture presents two major

drawbacks with the regard to the network performance.
First, there is no direct communication between the access
points and hence all traffic - external and intra-ring - has Fig.2: The phenomenon of bandwidth fragmentation
to pass through the hub. Next, edge nodes cannot receive
or remove transit frames on upstream channels (i.e., no As illustrated in Fig.2, the data channel can be in one of
spatial reuse of the bandwidth) or insert traffic on two states: free (idle) or occupied (busy). The first node on
downstream channels. upstream bus can fragment bandwidth in useless idle
periods. When the size of voids created at the first node is
This architecture implies multipoint-to-point not sufficiently large to allow the collision-free
communication on upstream bus and point-to-multipoint transmission of the item situated in front of the second
on downstream bus. We notice that the fairness issue queue, the transmission is blocked. Because the service
arises only on upstream bus, where several nodes compete discipline is based on FIFO scheduling, packets smaller
for the bus bandwidth. Since the hub node exclusively than one situated in head of the queue cannot use these
transmits on downstream wavelengths, traditional idle periods - the so-called Head of Line (HoL) blocking
scheduling mechanisms can be applied here. Therefore, phenomenon and, thus, the bus bandwidth is wasted
the access mechanism has to be controlled at the electronic
edge on upstream bus and this is the topic of the following
sections. 3.1 Positional Priority and Bandwidth Fragmentation
In what follows, we analyze the performance of the
3. The Performance Issue asynchronous CSMA/CA protocol. We consider a ring
with N nodes sharing a common medium (e.g., one optical
In this section, we firstly explain the positional priority channel) which channels traffic to the hub. The aim of this
and bandwidth fragmentation problems. Afterwards, we study is to present the average access delay (E[Wi}], i =
analyze the efficiency of the access protocol and highlight 1,..., N) and the distribution of the size of idle periods
the impact of the aforementioned issues on the network among different access nodes. We define access delay as
performance. the time spent by a packet in queue i until its successful
transmission.
To begin with, let us consider the situation when two
nodes share the bus. The channel is always idle for the In this study, simulations are carried using Network
first node on upstream bus. The packet situated in the Simulator 2 (ns-2) [9]. In our simulation works, unless is
head-of-line of the second node can access the available otherwise specified, we assume that:
bandwidth only if the transmission medium is free. ƒ All the ring nodes share a common upstream
Therefore, the transmission process of the second node optical channel with a capacity of 10 Gbps.
200 IJCSNS International Journal of Computer Science and Network Security, VOL.6 No.8A, August 2006

ƒ We assume a number of eight nodes (i.e., N = 8) number six and eight. We assume an upstream bus loaded
competing for the same resource. at 80 percent of its capacity. For this analysis, an unit
ƒ The process of client arrivals is modeled by a represents 64 Bytes. Therefore, the transmission of the
Poisson process. largest optical frame (i.e., network MTU) requires a void
ƒ The arrival rate (e.g., λ) of clients at each edge with a length of approximately 25 units (i.e.,
node is identical (i.e., balanced load: λi = λj, i≠j, i,j= ⎡(1518 + 56) / 64 ⎤ = 25 units). From these results it is
1…N ) in order to underline the fairness issue. immediately evident that bandwidth fragmentation leads
ƒ The capacity of the electronic memory at each ring to fairness problems at downstream nodes: Bandwidth is
node is set to 5 Mbytes. arbitrarily fragmented and most of “sensed” idle periods
ƒ We consider an optical overhead of 56 Bytes, as are less than the largest optical frame.
described in [8].
The simulation results of average access delay are
For the rest of this article, ρi denotes the average traffic illustrated in Fig. 4. Results presented here are significant.
load to be guided by the upstream bus from the ith node to We observe that for light loads (i.e., ρ = 0.3), the protocol
the hub. We call ρ the average traffic load of the upstream performance is good. For moderate-to-heavy load (curves
N
bus ( ρ = ∑ ρ ). In addition, we consider electronic named “rho = 0.7” and “rho = 0.8”), the downstream
i
i =1 node's performance is strongly affected. The poor
packets of variable size (i.e., 64, 576 and 1518 bytes) performance of the access protocol is not due to the
representative of the peaks in packet size distribution in saturation of transmission media. The fairness problems
Internet. The packet size distribution is as follows [8]: 10 are important even the channel occupancy is not high.
percent of traffic volume is composed of packets of 64
bytes (i.e., TCP ACK), 40 percent of traffic volume is
composed of packets of 576 bytes and the remaining 50
percent represents packets of 1518 bytes (i.e., Ethernet
MTU).

ρ = 0.8

Fig. 4: Access delay versus position on ring

In Fig. 5, we depict the channel utilization as a function of

the offered load (i.e., ρ). Note that the curve named
“Ideal” represents the ideal ratio of bandwidth utilization
Fig. 3: Statistics on void lengths when no bit loss occurs and no optical overhead is used to
transport client layer information. On one hand, we
Before presenting the simulation results, notice that 95 observe that network remains stable for light loads of the
percent confidence intervals were computed. Because the bus; a threshold of ρ = 0.6 can be defined. Beyond this
obtained width of confidence intervals is of 10-3, the threshold the channel utilization becomes saturated. This
confidence intervals are not drawn here in order to not has a simple explanation. Because most of the voids
obscure the data. “sensed” at downstream nodes are less than network MTU,
the process of transmission is blocked once the largest
In Fig. 3, we present statistics on idle periods circulating electronic packet arrives in head of downstream queues.
on bus. The curves plotted here illustrate the void length Packets smaller than one situated in front-of-queue cannot
distribution as measured at transmission tap of the node use these idle periods and, thus, the buffer overflow occurs.
IJCSNS International Journal of Computer Science and Network Security, VOL.6 No.8A, August 2006 201

On the other hand, we remark that the channel utilization,

as measured by simulation, increases faster than the Generally, fairness issue in ring architecture can be solved
“Ideal” case. This is due to the optical overhead used to following two approaches. The former introduces waiting
transport electronic packets at optical level. For example, lines in ring nodes to intercept the traffic in transit (i.e.,
the transmission of a small electronic packet of 64 Bytes RPR [4]). This approach does not fit to the DBORN
creates an overload (also called tax) of approximately [1- architecture because ring nodes are transparent and cannot
(64/(64+56))]*100 = 50 percent. process the transit traffic. Yet, introducing in-line buffers
at each node increases the network costs and this is an
important constraint in networking in metropolitan area
[11]. As for the later, we can proceed as follows. The
unfairness is dynamically detected and explicitly signaled
to upstream nodes (e.g., ATMR [15], MetaRing [14]).
Otherwise, ring nodes limit locally their activity using a
transmission-based quota (i.e., Token Bucket - TB).
Explicit signaling of unfairness problem relies on the
detection of node(s) which cause(s) starvation. On one
hand, such approach fails in DBORN because the network
throughput can reduce significantly due to the important
time of distribution of the control message between
neighbor nodes (i.e., a message makes a complete tour of
the ring to pass to both downstream and upstream stations).
On the other hand, the access protocol detects only the
presence or not of the light on transmission media and,
thus, cannot extract information from the transit traffic.

Fig. 5: Channel utilization vs. offered load (ρ) Local fairness algorithms using transmission-based quota
can share the bandwidth fairly among ring nodes. Yet, the
The results presented in this section show clearly that the bandwidth access at ring nodes remains arbitrarily.
phenomenon of bandwidth fragmentation leads to Bouabdallah et al. [2] have shown that transmission-based
significant performance degradation at downstream nodes. quota mechanisms (i.e., TB rate-based) do not ensure
Here, the mismatch between the size of idle periods satisfying results in asynchronous-based architecture like
circulating on the transmission medium and the length of DBORN. In addition, they analyzed the behavior of the
frames situated in head of downstream queues relies in TCARD scheme and showed that TCARD provides a
bandwidth waste as well as fairness problems with regard somewhat higher throughput to the access protocol.
to resource access. However, TCARD leads to unexpected behavior at heavy
traffic loads, as we shall see next.
To give an end, we underline that fairness algorithms to be
adopted in shared-bus architecture should treat equally 4.1 TCARD Limitations
both bandwidth fragmentation and positional priority
fairness-related issues. As we shall explain in the next In what follows, we provide a short description of
section, most of traditional fairness algorithms do not fit TCARD mechanism. For further details, the reader is
with DBORN architecture, since others cannot deal invited to refer to [6].The basic operation of such scheme
efficiently with the phenomenon of bandwidth is as follows. For each ring node i (i = 1,..., N-1), TCARD
fragmentation. maintains an anti-token pool. A TCARD anti-token
prevents the node i from transmission, and hence reserve
bandwidth, for a fixed amount of time d, on a detected idle
4. Related Works in Fairness Issue period. The size of d is upper bounded by the length of the
FDL (i.e., the length of network MTU). Idle periods
“sensed” on the bus can be used by the node for its own
Between the algorithms proposed to deal with fairness usage purpose as long as the anti-token pool is empty. The
issue in ring topologies, many of them require operations arrival of an anti-token to the pool during the transmission
that are not well suited to very high-speed optical of a frame does not preempt the emission process of a
networks (e.g., DQDB [16]) or are incompatible with such local frame. The anti-token has to wait the end of packet
dual-bus ring architecture (e.g., S++ [13]} requires transmission in order to be served. The rate of anti-tokens
operation at bit-level on transit traffic).
202 IJCSNS International Journal of Computer Science and Network Security, VOL.6 No.8A, August 2006

(i.e., B) and the size of the anti-token are two parameters on upstream bus fragments arbitrarily the bus bandwidth
directly related to the TCARD scheme. The anti-token rate between two anti-tokens.
Bi, at node i, corresponds to the total amount of average
bandwidth to be reserved for its (N-i) downstream partners,
and can be computed according to the following formula:

N
C∗ ∑ρ
j =i +1
j

Bi = (1)
d

Here, C denotes the channel capacity, ρj is the portion of

bandwidth reserved for the node j, and d is the size of the
anti-token.

Fig. 6 plots the void length distribution as “seen” by the

fourth node, as TCARD is enabled and disabled
respectively. For this study we consider a bus loaded at 70
percent (i.e., ρ = 0.7) of its capacity. We assume d =
network MTU = (1518+56) = 1574 Bytes. As for
parameter Bi, if the first node must reserve 6.125 Gb/s on Fig. 7: Access delay: CSMA/CA vs. TCARD
average of available bandwidth for downstream nodes,
anti-tokens are generated periodically at rate equal to Even if TCARD preserves a certain quantity of voids
(6.125 Gb/s)/(1574*8 b/anti-token) anti-tokens/s. Similar equal to network MTU, the starvation can still take place
to the first node, the second node reserves 5.25 Gbp/s for at relatively heavy traffic loads. The curves depicted in Fig.
its downstream partners, and so on. The curves show 7 come in favor of the previous affirmation. The results
clearly that the phenomenon of bandwidth fragmentation show the average access delay for an upstream bus loaded
is not significantly alleviated. at 70 and 80 percent respectively of its capacity. When ρ =
0.7, the mean waiting time grows fast in the beginning and
then tends to be constant or even decreases. This implies a
better fairness than the case when TCARD is disabled
(curve called “CSMA/CA: rho = 0.7”). With ρ = 0.8 and
TCARD enabled, the curve "explodes", then tends to be
constant and sinks for the downstream nodes. This leads to
unfairness: The starvation moves backward to upstream
nodes. Such unexpected behavior can be explained by two
aspects: i) the amount of large free periods sensed by an
edge node is not enough to consume immediately anti-
tokens and, ii) frames transmission is stopped until the
anti-token pool is emptied.

Certainly, when the TCARD algorithm is enabled,

upstream nodes reserve bandwidth for their downstream
partners. However, the size of voids reserved for
downstream nodes is not sufficiently large to avoid
starvation for moderate-to-heavy load. In this context, we
Fig. 6: Statistics on void length: The cdf measured at transmission tap of introduce a fairness algorithm that aims controlling
the 4th node; ρ = 0.7 packets transmission smarter than TCARD.

TCARD has a major drawback. The manner of bandwidth

reservation cannot substantially avoid the phenomenon of 5. The Proposed Algorithm: SPACES
bandwidth fragmentation. On one hand, downstream
nodes can fragment arbitrarily the idle periods reserved by Let us consider N nodes sharing a common unidirectional
their upstream partners. On the other hand, the first node channel. For each ring node i (i = 1, ..., N), SPACES
maintains two timers (i.e., reservation Treserv and
IJCSNS International Journal of Computer Science and Network Security, VOL.6 No.8A, August 2006 203

transmission Ttransm timers respectively). These timers are lengths several degrees of magnitude bigger than network
used to constraint the access to the transmission medium MTU.
of the ith node.

Fig. 9: Bandwidth fragmentation: CSMA/CA vs. SPACES

To illustrate the efficiency of the SPACES scheme, we

present in Fig. 9 a simple example with two-nodes sharing
Fig. 8: SPACES algorithm – State transition diagram a common bus. This picture shows immediately the main
advantage of the SPACES algorithm. Bandwidth
The basic operation of such a scheme is simple. As shown fragmentation is significantly alleviated. When SPACES is
in Fig. 8, our fairness algorithm has three states: Idle, enabled, the first node creates only large idle periods.
Restricted Access (RA) and Free Access (FA), respectively. These idle periods can be exploited by a downstream node
Initially, the SPACES algorithm is in Idle state. Packets to further preserve bandwidth for its downstream partners,
arriving to an empty buffer trigger the reservation timer transmit a burst of a size of Ttransm or emit a burst of
and determine the transition to the RA state. In this state, smaller size.
nodes cannot transmit data as long as the reservation timer
is not fired. The motivation for buffering packets during Unfortunately, however, the phenomenon of bandwidth
Treserv time units is to minimize the phenomenon of fragmentation cannot be completely avoided. Because the
bandwidth fragmentation and to create idle periods of process of transmission remains asynchronous, bandwidth
Treserv time units. When the reservation timer is fired, the is fragmented when a node on the ring does not exploit the
protocol switches from the state RA to the state FA, and totality of an idle period. Therefore, the mismatch between
the transmission is released for a period of Ttransm time the fixed-size of reserved idle periods and the variable
units. In this state, nodes can transmit only a quota of data length of optical frames will always lead with bandwidth
units before they transit to either Idle or RA state. If the fragmentation.
buffer is emptied during the FA state, timers are stopped
and the protocol transits to Idle state. When the end of The proposed method has two obvious advantages. First,
Ttransm occurs and the buffer is not empty, the protocol SPACES can be a distributed fairness algorithm using
transits back to the RA state. The end of Ttransm does not transmission-based quota: The average rate of traffic
preempt the transmission of a frame in order to keep the transmitted by a ring node is equal to its sustainable rate
complexity low and to avoid wasting bandwidth with specified in Service Level Specification (SLS) (e.g., ri).
packets partially transmitted. Yet, the current packet in Each ring node limits its transmission by setting properly
emission must be buffered for a latter transmission. the values of Treserv and Ttransm. For example, a node i
Therefore, the SPACES mechanism has to wait the access a ratio of channel capacity described by the
completion of the service of current frame to stop following formula:
transmission.
i
Ttransm 1 1
The SPACES algorithm has two key features. First, it ri = = = (2)
alleviates the positional priority problem by forcing ring
i
Ttransm + Treserv
i
T i
1 + ai
1 + reserv
i
nodes to smooth smartly the transmission of local traffic: Ttransm
Each ring node controls its transmission by a series of
busy and idle periods of activity. Next, unlike TCARD
The flexibility of the scheme permits to strictly limit the
which reserves voids upper bounded to the size of the
average rate of traffic transmission at upstream nodes
FDL, the SPACES algorithm can create idle periods with
while creating large idle periods. SPACES can also be a
204 IJCSNS International Journal of Computer Science and Network Security, VOL.6 No.8A, August 2006

statistical and fully distributed fairness algorithm. In this the hub at a mean rate of 1 Gb/s. Thus, the upstream bus is
case, the average rate of traffic transmitted by each ring loaded at 80 percent of its capacity.
node is no more limited by the sustainable rate specified in
SLS. This is a significant difference of SPACES with
respect to TB and TCARD respectively. For example, if
the first node uses only a small portion of available
bandwidth, the second node can benefit from the unused
part of the first node. The third node can access the unused
bandwidth of the previous two nodes and so on.
Conversely, TB and TCARD limit the transmission of
each ring node to its sustainable bit rate independent of the
current state of the resources usage.

The statistical version of the SPACES scheme must deal

with the problem of finding an optimum solution for
setting of the values of Tireserv and Titransm respectively.
When ai < 1 (i.e., Tireserv < Titransm) , a node grabs more
bandwidth than it reserves for its downstream partners.
This situation leads to unfairness with the regard to access
bandwidth at downstream nodes. When ai > 1, bandwidth
is wasted if upstream nodes reserve a ratio of bus capacity Fig. 10: Bandwidth fragmentation as “seen” by the fourth node:
larger than the load of the traffic to be transmitted at SPACES vs. TCARD & CSMA/CA
downstream nodes. The case ai = 1 leads to what we call
local fairness with the regard to downstream partners: A Fig. 10 depicts the interaction between the aforementioned
ring node transmits Titransm time units and, afterwards, schemes and the phenomenon of bandwidth fragmentation.
stops its transmission for an identical period of time. We assume Tireserv = Titransm = 50 μs, i = 1…N. The results
Notice that at channel fully loaded, the upstream nodes confirm the efficiency of the SPACES algorithm: We
create a series of identical busy and free periods. observe that the ratio of large-size idle periods increases
significantly.
We highlight that finding an optimum solution for setting
of values of Tireserv and Titransm respectively is not an easy
task. However, when local fairness version of the
SPACES algorithm is used, the network performance
increases significantly, as we shall show in the next
section.

6. Performance Evaluation of SPACES

In this section, we evaluate the performance of the
SPACES scheme and compare it with the one achieved by
TCARD. We notice that we keep similar assumptions used
within the simulation works presented earlier. For
SPACES, we consider only the case of local fairness (i.e.
a = 1) for a main reason: The parameters setting of
SPACES is independent of the physical location of a node
within a network. The parameters setting in TCARD is as
defined in the previous section.
Fig. 11: Interaction between the SPACES scheme and bandwidth
fragmentation
We first compare the performance of each ring node when
the SPACES algorithm is enabled, when the TCARD The probability of “sensing” large free periods at end of
algorithm is enabled, and when both SPACES and the bus increases as the size of the reservation timer is
TCARD algorithms are disabled. Each node receives increasing. To get a more detailed insight into the
traffic from the access networks to be forwarded towards correlation of the size of reservation and transmission
IJCSNS International Journal of Computer Science and Network Security, VOL.6 No.8A, August 2006 205

timers and the protocol performance, on the other hand, optical/physical level, as it was explained in Section 3. For
we consider three cases: A) Tireserv = Titransm = 50 μs, B) loads less than 1.0, the efficiency of the SPACES
Tireserv = Titransm = 250 μs, and C) Tireserv = Titransm = 500 μs, algorithm is evident: The system remains stable up to a
i = 1…N. load of ρ = 0.9. For ρ = 0.9, we notice a light performance
degradation at the last node on the upstream bus (i.e., eight
The curves plotted in Fig. 11 show clearly that an increase node).
in reservation timer leads to an increase in length of voids
“sensed” at downstream stations. Here, we present the
distribution of the size of idle periods, as “seen” at the
transmission tap of the fourth node.

SPACES adds an additional delay of Treserv time units at

upstream queueing times. Results, in terms of average
access delay, depicted in Fig. 12 show that the fairness
issue is alleviated at downstream nodes at expense of
media access delay increasing at upstream nodes. We
observe that in all three cases the relative performance
between the ring nodes remains fair. It is clear that the
efficiency of the SPACES algorithm in achieving fairness
is independent of its parameters setting. We notice that no
data loss is observed at ring nodes when the SPACES
algorithm is enabled. We point out, however, that larger
the reservation window larger access delays. The delay
recorded for upstream nodes is increased, compared to the
case where no fairness control mechanism is applied or Fig. 13: Access delay for heavily loaded channel
TCARD is enabled, but still remains in acceptable limits.

Fig. 14: Node throughput for heavily loaded channel

Fig. 12: Access delay as a function of position on bus: SPACES versus
TCARD & CSMA/CA
Of course, the system performance is nothing but an
In Fig. 13 and 14, we show SPACES behavior, in terms of answer to a dedicated load. Therefore, results would differ
average access delay and node throughput respectively, from the curves shown in this paper if different arrival
when the bus is heavily loaded (i.e., ρ = 0.9 and 1.0). We process would be assumed. However, our goal was to
observe that the SPACES algorithm may be unfair in the prove that the distributed version of the SPACES
sense that the senders situated at the end of the bus are algorithm avoids efficiently the phenomenon of bandwidth
saturated when the system is 100 percent loaded. This is fragmentation while achieving fairness among ring nodes.
due to the phenomenon of bandwidth fragmentation and In addition, simulation results depicted in Fig. 15 show
the low efficiency of transporting electronic packets at that SPACES improves significantly the resource
206 IJCSNS International Journal of Computer Science and Network Security, VOL.6 No.8A, August 2006

utilization: The network remains stable for loads up to 0.8. traditional fairness algorithms: It is fully distributed. It
Thus, the ratio of bandwidth utilization increases from 60 means that the parameters setting of SPACES is
to 80 percent. independent of the physical location of a node within the
network. Simulation results showed that the proposed
solution alleviates substantially the performance
SPACES degradation at heavy bus loads while achieving fairness
among ring nodes.

TCARD The results on which the comparison was based provide

many useful insights. For moderate-to-high load, the
CSMA/CA phenomenon of bandwidth fragmentation can be
substantially alleviated if the upstream nodes create large
fixed-size idle periods. However, the mismatch between
the fixed-size of the reserved voids and the variable-length
of frames will always lead to bandwidth fragmentation in
asynchronous CSMA/CA scheme. This means that little
can be done to achieve throughput and access delay
fairness when the traffic load is heavy.

There are other practical aspects that have to be studied. In

this paper, we have not addressed the ability of the
Fig. 15: Channel utilization versus offered load (ρ) SPACES scheme to deal with traffic with different QoS
requirements. As described in this article, SPACES uses
two timers (i.e., reservation and transmission timer
respectively) to regulate the process of transmission. The
7. Conclusions reservation timer is a key parameter for the size of idle
periods “sensed” at downstream nodes, and hence for
In this paper, we have studied the fairness issue of the downstream access delays, as well as for access delays at
asynchronous optical CSMA/CA protocol with variable upstream nodes. A “target” access delay is strictly
packet sizes. For light-to-moderate load, the access guaranteed at upstream nodes and only “statistically”
protocol provides good performance while achieving achieved at downstream nodes. However, it is expected
fairness among ring nodes. The fairness discussion for that queueing times at downstream nodes will decrease as
heavy load has shown that the unfairness, in regard with the size of the spectral dimension (i.e., the number of
the access of the shared bandwidth, is pronounced upstream wavelengths) is increasing.
between upstream and downstream nodes.

The asynchronous CSMA/CA scheme suffers from References

bandwidth fragmentation and positional priority fairness-
related problems. First, we proved the former represents [1] J. Tomasik and D. Popa, “On Markov Chain Modelling for
Asynchronous Optical CSMA/CA Protocol”, submitted to
the most important issue, with the regard to the access
IEEE Communications Letters.
protocol performance, because it increases significantly [2] N. Bouabdallah, A.-L. Beylot, E. Dotaro and G. Pujolle,
the unfairness for downstream nodes. Next, we explained “Resolving The Fairness Issue in Bus-Based Optical Access
why the traditional fairness algorithms fail in achieving Networks”, IEEE JSAC, vol. 23, no. 8, pp. 1444-57, Aug.
satisfactory performance in such shared-bus configuration. 2005.
Finally, we exhibited the limitations of a previous [3] A. Carena, V. De Feo, J.M. Gaudino, R. Neri, C. Piglione
proposed fairness algorithm, the so-called TCARD. For and P. Poggiolini, “RINGO: An Experimental WDM Optical
moderate-to-heavy load, we observed that TCARD cannot Packet Network for Metro Applications”, IEEE JSAC, vol.
deal efficiently with the fairness issue in DBORN. 22, no. 8, pp. 1561-71, Oct. 2004.
[4] F. Davik et al., “IEEE 802.17 Resilient Packet Ring Tutorial”,
IEEE Commun. Mag", vol. 42, no. 3, pp. 112-18, Mar. 2004.
Consequently, we considered an additional flow control
[5] I.M. White, M.S. Rogge, K. Shrikhande and L.G. Kazovsky,
mechanism not only to limit the node transmission but also “A summary of the HORNET project: A next-generation
to regulate the emission process on shared bandwidth. We metropolitan area network”, IEEE JSAC, vol. 21, no. 9, pp.
presented a statistical fairness algorithm, called SPACES, 1478-94, Nov. 2003.
which contains a major advantage over TCARD and
IJCSNS International Journal of Computer Science and Network Security, VOL.6 No.8A, August 2006 207

[6] L.Ciavaglia, N.Bouabdallah, E. Dotaro and N. Le Sauze, in access networks. His recent research interests include optical
“Matching fairness and performance by preventive traffic MAN/WAN networks and wireless access networks. He was
control in optical multiple access networks”, in Proc. Of involved in many French national research projects (ROM,
OptiComm2003, pp. 424-429, Oct. 13-17 2003, Dallas, USA ROM-EO, IPSAT), as well as the European Network of
[7] E. Dotaro et al., “DBORN: A Shared WDM Ethernet Bus Excellence.
Architecture for Optical Packet Metropolitan Network”, in
Proceedings of Photonics in Switching PS'02, July 2002. Tülin ATMACA received the PhD
[8] “IP Packet Length Distributions”, degree in Computer Science from The
http://www.caida.org/analysis/AIX/plein\_hist, Juin 2002. University of Paris XI (ORSAY), in
[9] “Network Simulator", http://www.isi.edu/nsnam/ns, 2002. France in 1987. From 1986 to 1988 she
[10] N. Le Sauze et al., “A Novel, Low Cost Optical Packet has taught in the Department of Computer
Metropolitan Ring Architecture”, in Proceedings of Science at the University of Paris XI.
ECOC’01, pp. 66-67, 30 Sept. - 4 Oct. 2001. From 1989 to 1991, she was Assistant
[11] R.H. Cardwell, O.J. Wasem and H. Kobrinski, “WDM Professor in the Computer Science
Architecture and Economics in Metropolitan Area”, SPIE Department at the University of Paris VI.
Optical Network Mag., July 2000. Since January 1992, she has been Associate Professor at National
[12] M.J. O'Mahony, “The Application of Optical Packet Institute of Telecommunications (INT) in France, in the
Switching in Future Communication Networks”, IEEE networking and service department.
Comm. Mag. pp. 128-25, Mar. 2001. She was a Visiting Professor in the Computer Science
[13] G.C. Watson and S. Tohme, “S++ - A New MAC protocol Department of the North Carolina State University, as well as the
for Gb/s Local Area Networks”, IEEE JSAC, pp. 531-539, department of Industrial Engineering of the Rutgers University.
vol.41, no. 4, May 1993. Her research interests are in the areas of performance evaluation
[14] I. Cidon and Y. Ofek, “MetaRing - A Full-Duplex Ring with of telecommunication networks (ATM, FR, Optical Networks,
Fairness and Spatial Reuse”, IEEE Trans. on Comm., vol. 41, TCP/IP,...), of traffic and congestion control, and Quality of
no. 1, pp. 110-120, Jan. 1993. Services aspects in these networks. She is involved in several
[15] T. Okada, H. Ohnishi and N. Morita, “Traffic Control in national and international research projects in the field of the
Asynchronous Transfer Mode”, IEEE Comm. Mag., pp. 58- optical packet switching network and their performance (ROM,
62, vol. 29, no. 9, Sept. 1991. ROM-EO, EuroNGI, Industrial research projects). She served on
[16] "Distributed Queue Dual Bus (DOQB) Subnetwork of a the program committees and was co-chair of various well-known
Metropolitan Area Network", IEEE 802.6 WG, Ed. international conferences (AICT’06, ICDT’06). Her recent
P802.6/D15, Oct. 1990. research interests include the all-optical multi-service networks:
[17] A.A. Lazar, A.T. Temple and R. Gidron, “MAGNET II: A WAN and MAN and the performance of optical access networks.
metropolitan are network based on asynchronous time
sharing”, IEEE JSAC, vol. 8, no. 8, pp. 1582-94, Oct. 1990.
[18] R.M. Falconer and J.L. Adams, “Orwell: A protocol for an
integrated service local network”, British Telecom. Technol.
J., vol. 3, pp. 27-35, no. 4, Oct. 1985

Daniel POPA received the M.Sc. degree

in telecommunication engineering from the
Polytechnic University of Bucharest,
Romania, in 2001, and the PhD degree in
Computer Science from National Institute of
Telecommunications, France, in June 2005.
He is currently working in the Optical
Network Division of ALCATEL R&I, in
France, where he is Research Engineer in
the Packet Transport Infrastructure Laboratory.
He was a Visiting Researcher in the Computer Science
Department of Academy of Mathematical Sciences, Poland, in
July 2002 and June 2003, as well as the Electrical Engineering
Department of Istanbul Technical University, Turkey, in October
2004. From June 2005 to August 2006, he held a Research
Fellowship in the Computer Science Department of National
Institute of Telecommunications, France.
He has done extensive research in areas of performance
evaluation, traffic management, Quality of Services and fairness