Computer Networks xxx (2015) xxx–xxx

Contents lists available at ScienceDirect

Computer Networks
journal homepage: www.elsevier.com/locate/comnet

On efficiency and validity of previous Homeplug MAC

performance analysis
C. Cano ⇑, D. Malone
Hamilton Institute, National University of Ireland, Maynooth, Co. Kildare, Ireland

a r t i c l e i n f o a b s t r a c t

Article history: The Medium Access Control protocol of Power Line Communication networks (defined in
Received 5 February 2015 Homeplug and IEEE 1901 standards) has received relatively modest attention from the
Accepted 5 March 2015 research community. As a consequence, there is only one analytical model that complies
Available online xxxx
with the standardised MAC procedures and considers unsaturated conditions. We identify
two important limitations of the existing analytical model: high computational expense
Keywords: and predicted results just prior to the predicted saturation point do not correspond to
Power Line Communications
long-term network performance. In this work, we present a simplification of the previously
Homeplug
IEEE 1901
defined analytical model of Homeplug MAC able to substantially reduce its complexity and
Mean field analysis demonstrate that the previous performance results just before predicted saturation corre-
Decoupling approximation spond to a transitory phase. We determine that the causes of previous misprediction are
Priority resolution scheme common analytical assumptions and the potential occurrence of a transitory phase, that
we show to be of extremely long duration under certain circumstances. We also provide
techniques, both analytical and experimental, to correctly predict long-term behaviour
and analyse the effect of specific Homeplug/IEEE 1901 features on the magnitude of mis-
prediction errors. Our simplified analysis is then used to model the priority resolution
scheme defined in Homeplug/IEEE 1901. We show that previous tendencies found in sim-
ulations and in an off-the-shelf testbed are correctly captured in our framework. We also
show that resorting to the traffic differentiation provided by the different contention
parameters alone solves the starvation faced by lower-access category traffic.
Ó 2015 Elsevier B.V. All rights reserved.

1. Introduction
Data transmission using electrical wires, known as
Power Line Communication (PLC), has recently experi-
enced increased deployment as a useful complement and
strong competitor to wireless networking solutions. For
instance, the chip manufacturer Qualcomm Atheros
reports over 100M Homeplug networking devices shipped
and expects this number to grow by 32% each year from
2011 to 2017 [1]. The appeal of PLC relies in the high data
⇑ Corresponding author.
E-mail addresses: cristina.cano@nuim.ie (C. Cano), david.malone@
nuim.ie (D. Malone).
rates it can deliver, its low deployment cost (as it is easy to
retrofit) and allows communication through obstacles that
commonly degrades wireless signals. Additionally, it also
provides a low-cost alternative to complement existing
technologies to reach ubiquitous coverage. For instance,
as a backhaul for wireless sensor networks or small cells.
Research efforts have been mostly focused on the physi-
cal layer as the characteristics of PLC channels (including
fading, impulsive noise and hidden/exposed terminal prob-
lems) impose several challenges on physical aspects [2].
However, although some issues related to the Medium
Access Control (MAC) defined in Homeplug [3] and IEEE
1901 [4] standards which may severely affect user satisfac-
tion have been identified (such as unfairness [5] and

http://dx.doi.org/10.1016/j.comnet.2015.03.005
1389-1286/Ó 2015 Elsevier B.V. All rights reserved.

Please cite this article in press as: C. Cano, D. Malone, On efficiency and validity of previous Homeplug MAC performance analysis, Comput.
Netw. (2015), http://dx.doi.org/10.1016/j.comnet.2015.03.005
2 C. Cano, D. Malone / Computer Networks xxx (2015) xxx–xxx
starvation of lower access categories [6]), the MAC protocol
of these standards has not received much attention.
Homeplug [3] and IEEE 1901 [4] define a MAC proce-
dure similar to the Distributed Coordination Function
(DCF) defined in the IEEE 802.11 standard for Wireless
Local Area Networks (WLANs) [7]. PLC MAC protocols
diverge from the vanilla DCF by adding a deferral counter
that reduces the attempt rate when high contention is
inferred on the channel (i.e., a certain number of packets
are overheard). Despite being a substantial change over
DCF, this feature has not been deeply evaluated.
Attempts to compare DCF and PLC MAC protocols have
been made in [8,9,5,10]. However, there is still much work
to do in order to fully demonstrate the conditions under
which the deferral counter improves the performance of
the network.
Contributions to the analysis of PLC MAC protocols aim
to extend the research on performance evaluation of PLC
networks, crucial to advance understanding and optimisa-
tion. The only relatively complete analytical model of
Homeplug MAC is presented in [11]. This analytical model
has been widely used [12–17], however we identify two
significant issues. The first is the high computational com-
plexity which results in: (i) unsuitability for use at runtime
in order to tune parameters based on changing network
conditions and (ii) difficulty of extension to consider other
features of the standards that are implemented in current
off-the-shelf devices, such as the priority resolution
scheme, aggregation or buffer management mechanisms.
Second, as we will demonstrate, the performance results
obtained in the regime right before the predicted sat-
uration point do not correspond to the long-term network
behaviour. We also use our simplified framework to model
heterogeneous access categories and the strict priority res-
olution scheme defined in Homeplug and IEEE 1901. In
detail, the main contributions of this work are the
following:
 We propose a reformulation of the analytical model for
the Homeplug MAC procedure presented in [11] that
facilitates a simplified method of solution. Our pro-
posed analysis provides a 2 orders-of-magnitude
improvement in runtime compared to the one pre-
sented in [11] while maintaining its accuracy.
Specifically, we: (i) apply a renewal reward approach
thus avoiding the need to calculate a full stationary dis-
tribution for a Markov chain [18,19], (ii) allow the most
expensive operations to be precomputed and (iii) pro-
vide and evaluate the accuracy of an optional exponen-
tial approximation to the probability to defer in a given
backoff stage.
 We use our simplified analytical framework to extend
the analysis to heterogeneous access categories and to
model the strict priority resolution scheme defined in
Homeplug and IEEE 1901. This is the first work to pre-
sent an analytical model of Homeplug MAC and IEEE
1901 when different access categories contend for the
channel. Although simplicity is maintained as one of
the key aspects of our analysis also for this case, results
agree with simulations and capture the tendencies
observed in an off-the-shelf testbed, including
Please cite this article in press as: C. Cano, D. Malone, On efficiency and validity of previous Homeplug MAC performance analysis, Comput.
Netw. (2015), http://dx.doi.org/10.1016/j.comnet.2015.03.005
previously observed strong starvation of lower-access
categories under high traffic load [20,6]. Traffic dif-
ferentiation through contention access parameters
instead of by the use of the priority resolution scheme
is proposed as one practical and simple way to solve
the identified starvation problem.
 We demonstrate that the results right before the pre-
dicted saturation point presented in [11] correspond
to a transitory phase of the system instead to the
long-term behaviour. We identify the two causes of
misprediction in [11]: (i) the decoupling approximation
under infinite buffer size considered in the analytical
model and (ii) the presence, under certain circum-
stances, of an extremely long (of the order of magnitude
of hours) transitory phase in simulations. This is the
first work that shows the long duration of that transi-
tory phase and, indeed, the causes of its occurrence sug-
gest it may also be present in generic random access
protocols. Having identified the causes of misprediction,
we provide techniques to generate valid results from
analytical models which use the decoupling approx-
imation as well as from simulations.
 We evaluate the impact of the deferral counter on the
magnitude of misprediction errors by evaluating the
different solutions obtained considering: (i) the starting
values of the deferral counter proposed by the standard,
(ii) no deferring (as done in IEEE 802.11 DCF) and (iii)
always deferring after overhearing following the pro-
posal presented in [21].
The remainder of this article is organised as follows. In
Section 2, we review the Homeplug MAC procedure. Next,
in Section 3, related work on PLC MAC analytical models
is discussed. Our simplified analytical model, including
the extension to heterogeneous access categories, is
described in Section 4. Then, in Section 5, we demonstrate
that the previous performance results right before pre-
dicted saturation do not correspond to the long-term
behaviour of the network. The performance evaluation,
including the validation of the presented analytical
model, an evaluation of the effect of the deferral counter
on misprediction errors, performance results using het-
erogeneous access categories and a complexity compar-
ison, are presented in Section 6. Finally, some final
remarks are provided.
2. Homeplug MAC background
The Homeplug MAC protocol is similar to the DCF
defined in the IEEE 802.11 standard [7]. Each time a node
has a new packet to transmit, the backoff stage
(i 2 ½1; m)1 is initialised to 1 and a random backoff is
selected among ½0; W 1 . The backoff countdown is frozen
when activity is detected on the channel and restarted when
the medium becomes idle again. The packet is actually
transmitted when the backoff countdown expires. If an
acknowledgement is received, the packet is considered
1
Actually, (i 2 ½0; m  1) but indexes have been relabelled here for
clarity of illustration.
 C. Cano, D. Malone / Computer Networks xxx (2015) xxx–xxx
successfully transmitted. Otherwise, the node starts the
retransmission procedure: the backoff stage changes to
i ¼ minði þ 1; mÞ and a new random backoff is selected
among ½0; W i , being W i the contention window of stage i.
Additionally, in the Homeplug and IEEE 1901 MAC, a
new counter, called the Deferral Counter (DC), is intro-
duced. This counter is initialised at each backoff stage
to M i and decreased by one after overhearing a data
packet or a collision. If a new packet or a collision are
overheard and the value of the DC is equal to zero, the
node acts as if a collision had happened: the backoff stage
is increased if it has not yet reached its maximum value
and a new backoff is selected among ½0; W i . The goal of
the DC is to avoid collisions when high contention is
inferred by decreasing the aggressiveness of transmission
attempts.
Fig. 1 depicts the part of the Markov Chain correspond-
ing to a given backoff stage as modelled in [11], where each
state represents the current backoff stage, the value of the
deferral counter and the backoff countdown. The combined
behaviour of the backoff countdown and the DC counter is
shown. It can be observed how the DC counter is decreased
when the channel is detected busy (that happens with pb
probability) and how the new backoff stage is increased
when the DC counter reaches zero and a new packet or col-
lision are overheard.
To provide channel access differentiation, 4 CAs are
defined CA0–3. CA3 and CA2 share W i and M i values, as
do CA1 and CA0 (see Table 1). Two Priority Resolution
Slots (called PRS0 and PRS1) are allocated at the end of suc-
cessful frame exchanges as shown in Fig. 2. These slots
allow nodes to announce the priority of packets pending
transmission. The highest priority (CA3) is signalled by
transmitting a symbol in both PRS0 and PRS1, the CA2
category is signalled in PRS0 only. CA1 signals in PRS1, if
PRS0 was empty, and the lowest access category (CA0)
does not signal at all. Following this approach, stations
know if there is a station with a frame that belongs to a
higher CA. In such a case, they do not contend for the chan-
nel, allowing high-priority frames to be released.
Note that this resolution scheme aims to provide strict
access differentiation, i.e., using the priority resolution
mechanism, packets with higher priority are always trans-
mitted before lower-priority ones. However, the priority
resolution scheme is only invoked after successful frame
exchanges. The standards [3,4] suggest that PRS are not
present after: (i) a collision, (ii) frame transmissions result-
ing in erroneous receptions and (iii) the detection of an
empty channel for longer than an Extended InterFrame
Space (EIFS) period.2 Thus, in lightly loaded conditions and
after collisions or channel errors, the priority resolution
scheme is not employed and channel access differentiation
only occurs through the different parameters of the access
categories. Thus, we expect strict prioritisation if we have
a single station in a high CA, but less strict prioritisation if
multiple stations are in the highest CA because of collisions,
as shown in [20,6] via simulations and experimental evalua-
tion, respectively.
2
EIFS is set to the duration of a frame transmission of maximum length.
Please cite this article in press as: C. Cano, D. Malone, On efficiency and validity of previous Homeplug MAC performance analysis, Comput.
Netw. (2015), http://dx.doi.org/10.1016/j.comnet.2015.03.005
3
3. Related work on Homeplug MAC modelling
The analytical model in [11] takes into account all the
features of the Homeplug MAC procedure except the chan-
nel access prioritisation. It models the node access proce-
dure as a 3-dimensional Markov Chain in which the
backoff and deferral counters, as well as the different back-
off stages are considered. To solve for the stationary proba-
bilities of the Markov Chain an iterative numerical method
is required. Then, to compute the channel access delay, the
authors follow a recursive approach, i.e., the delay at every
backoff stage depends on the delay of previous stages. To
obtain this performance metric, further iterations are
needed. Moreover, to compute the channel access proba-
bilities as well as the delay, many computationally expen-
sive operations are required. The reason for this is the
geometric nature of the deferral counter expiration and
its dependence on the selected random backoff. The
authors also extend the model to consider unsaturated
conditions. However, the approach taken is based on fur-
ther iterations in which the value of the transmission
attempt rate is reweighted based on the offered load at
every iteration until a reasonable approximation is
obtained from the saturated model. As we will see in
Section 6, the cost of solving this model makes it pro-
hibitive for situations where results are required in real-
time.
A similar approach to ours which also uses the renewal
reward approach and that considers the backoff and defer-
ral counters has been presented in [22]. However, only
saturated conditions and homogeneous access categories
are considered. Recently, in [23], the same authors identify
a different inaccuracy arising from the decoupling approx-
imation under saturation conditions and a small number of
contending stations, due to the high coupling between
queues when the deferral counter is used. They also pro-
pose a model of the coupled system of queues based on
the saturated assumption. Note that our work differs from
[23] in the following ways: (i) we evaluate the implications
of the decoupling approximation on the buffers’ occu-
pancy, which can result in two different behaviours (see
Section 5) and (ii) we propose techniques to obtain the
long-term performance of the network when using a
model based on the decoupling approximation under
unsaturated conditions. We note that directly extending
the analysis of the coupled queues to nonsaturated condi-
tions appears intractable.
Two other analytical models of Homeplug exist in the
literature ([24,21]). However, they model mechanisms that
differ from the original standardised Homeplug procedure.
In [24], the value of the contention window is fixed for all
backoff stages, while in [21] the backoff stage is incre-
mented every time a new packet is overheard (thus having
M i ¼ 0; 8i). These new approaches obviously lead to a sim-
plified analysis, as one of the dimensions of the Markov
Chain is removed. However, the flexibility the standard
provides by allowing to tune both the W i and Mi is not cap-
tured by these frameworks.
Consequently, the analytical model in [11] has been
widely used both to derive performance metrics and as a
basis for extension [12–17] as it strictly follows the
4 C. Cano, D. Malone / Computer Networks xxx (2015) xxx–xxx
Fig. 1. Markov chain model in backoff stage i of the Homeplug MAC [11].
Table 1
Parameters of the different access categories in Homeplug.
Parameter All CAs Parameter CA3/2
M1 0 W1 7 7
M2 1 W2 15
M3 3 W3 15
M4 15 W4 31
procedure defined in the standard and it considers satu-
rated as well as unsaturated conditions. This widespread
use motivates our improvements in terms of (i) reducing
complexity while maintaining the accuracy and retaining
all features defined in the standard [3] and (ii) understand-
ing the transitory nature of the results predicted before
saturation in [11].
4. Simplified Homeplug analytical model
In this section, we present the simplified analytical
model of the Homeplug MAC protocol. To this end we take
a renewal reward approach [18,19] motivated by the fact
that the attempt rate of a given node can be viewed as a
regenerative process. This approach allows us to compute
metrics of interest without the need to solve all state
probabilities of the Markov Chain embedded in the analy-
sis. Similar approaches have already been proposed to sim-
plify the derivation of analytical models of the IEEE 802.11
for WLANs, see for instance [25,26].
We also apply the decoupling approximation to model
the conditional (given that a packet is transmitted) colli-
sion probability independently of the backoff stage, as
also done in [11] and in a large number of IEEE 802.11
analytical models (see [27] for a complete survey on this
Please cite this article in press as: C. Cano, D. Malone, On efficiency and validity of previous Homeplug MAC performance analysis, Comput.
Netw. (2015), http://dx.doi.org/10.1016/j.comnet.2015.03.005
topic). Although this assumption may not be valid [28], it
allows us to simplify the analysis while still accurately
CA1/0 predicting the metrics of interest in a broad range of cir-
cumstances. The decoupling approximation is also used
15 to model the buffer occupancy probability after a trans-
31 mission as being independent of the backoff stage at
63 which the packet was transmitted. This assumption is
implicitly made in [11]. The accuracy of this assumption
will be discussed in detail in Section 5 as it is contribut-
ing factor to misprediction of results right before sat-
uration in [11].
The rest of assumptions and considerations taken into
account are: (i) an infinite, or large enough to be consid-
ered infinite, buffer size3 and retry limit, (ii) exponentially
distributed interarrival of packets, (iii) ideal channel condi-
tions (i.e., packets are always received correctly in the case
of no collision), (iv) contention among a single access cate-
gory (extension to heterogeneous access categories is pre-
sented in Section 4.4), and (v) all nodes are in mutual
coverage range, that is, all nodes can overhear each other’s
transmissions. These are the same assumptions considered
in [11] and in a number of IEEE 802.11 analytical frame-
works. On one hand they allow us to compare the accuracy
of our analysis and the one in [11] under the same condi-
tions. On the other hand, they are useful to assess and
understand the performance of the MAC procedure indepen-
dently of other factors that may have an influence in the
results in practice.
3
Although our framework can be easily extended to finite buffer size, for
instance, as done in [26], we have chosen to maintain the infinite buffer
assumption so as to compare with the results in [11]. It is worth pointing
out that, in practice, the large dimensions of current buffers can as well be
approximated by a infinite buffer size assumption without losing accuracy.
 C. Cano, D. Malone / Computer Networks xxx (2015) xxx–xxx 5

Fig. 2. Allocation of priority resolution slots (refer to [3,4]).

4.1. Renewal reward approach The term I in Eq. (5) accounts for the number of slots in
idle state (when there is no packet waiting in the queue for
Assuming an infinite buffer size, the mean queue occu- transmission) and is computed as the probability of having
pancy (q) of a node is derived by considering the time an empty queue over the probability of a packet arrival in a
needed to release a packet from the queue (X), called ser- slot. Considering an M/M/1 queue, we then compute I as in
vice time or MAC access delay, and the packet arrival rate Eq. (6).
from the network layer (k) as q ¼ minðkX; 1Þ. While k  
1q
depends on the application, the service time is computed I ¼ max ;0 ð6Þ
1  eka
as the sum of the following three components: (i) the total
average backoff duration until the successful frame trans- The average slot duration while the node is in backoff is
mission, (ii) the total time on average spent in transmitting derived depending on the type of slot that is overheard (Eq.
packets that result in a collision and (iii) the time spent (7)). A slot can be empty if no other node transmits (that
successfully transmitting the packet: occurs with pe probability) and, in such a case, its duration
X ¼ E½wa þ ðnt  1ÞT c þ T s ; ð1Þ is r. Otherwise, it can be occupied due to a successful
transmission (that happens with probability ps ) or a colli-
where E½w denotes the average number of slots in backoff, sion (that occurs with pc probability), with durations T s
a is the average slot duration and nt is the average number and T c , respectively.
of attempts to successfully transmit a packet. The duration
of a successful transmission (T s ) and a collision (T c ) are
a ¼ ps T s þ pc T c þ pe r ð7Þ
computed as shown in Eq. (2). These durations account Probabilities ps , pe and pc are obtained as follows:
for the transmission notification intervals (PRS0 and
PRS1), the time to transmit the actual frame (T fra ) and ps ¼ ðn  1Þsð1  sÞn2 ;
the acknowledgement (T res ), as well as the contention pe ¼ ð1  sÞn1 ;
and response interframe spaces (CIFS and RIFS), see [3]. pc ¼ 1  ps  pe : ð8Þ
Note that considering that all nodes are in coverage range,
collisions involving acknowledgements are not possible. To compute X, it remains to obtain the value of the
However, their transmission time must be considered in average number of slots spent in backoff until a packet
a successful transmission.4 is successfully transmitted (E½w). With this aim, we fol-
low the same approach as in [11] but we compute the
T s ¼ T c ¼ PRS0 þ PRS1 þ T fra þ RIFS þ T res þ CIFS ð2Þ
average number of slots at every backoff stage. The aver-
Under the decoupling assumption with an infinite num- age number of slots waiting in backoff at backoff stage i
ber of retries, the average number of attempts to transmit a (with 1 6 i 6 m) is shown in Eq. (9). This expression is
frame (nt ) is computed as shown in Eq. (3). obtained from [11] but it has been rearranged to account
for the number of backoff slots instead of the total delay
1
nt ¼ ; ð3Þ to transmit a packet. The reader is referred to [11] for
1p
details.
where the conditional collision probability (p) is obtained "
as the complementary probability of having at least one 1 M2i þ M i
E½wi  ¼
of the other n  1 nodes transmitting a frame in the same Wi þ 1 2
slot (Eq. (4)), with s denoting the attempt rate of a node. ( !
WX
i M i Xk1 Mi þ l
M
p ¼ 1  ð1  sÞ n1
ð4Þ þ ð1  pb Þl pb i pb ðl þ M i þ 1Þ
k¼1 l¼0 l
We view the attempt rate as a regenerative process, ! )#
where the renewal events are when the MAC begins pro- Mi
X Mi þ k
M l
cessing a new frame. Thus, we apply the renewal reward þ ð1  pb Þkþl pb i ðk þ Mi Þ ð9Þ
l¼0 kþl
theorem (Eq. (5)).
nt In Eq. (9), pb refers to the probability of overhearing
s¼ ð5Þ another transmission while the node is in backoff:
E½w þ nt þ I
pb ¼ 1  ð1  sÞn1 : ð10Þ
4
In contrast to [11], we have considered T s ¼ T c to closely model the A node moves to the next backoff stage (if it has not
standard procedure. Observe that, in case of collision or frame errors, either
an NACK is received or nodes wait an EIFS to provide protection from
reached its maximum value) whenever it fails to transmit
collisions for ongoing transmissions [3]. Thus, the duration of a collision can a packet. We define a failure here as the case in which
be better approximated to that of a successful transmission. either: (i) the backoff expires leading to the transmission

Please cite this article in press as: C. Cano, D. Malone, On efficiency and validity of previous Homeplug MAC performance analysis, Comput.
Netw. (2015), http://dx.doi.org/10.1016/j.comnet.2015.03.005
6 C. Cano, D. Malone / Computer Networks xxx (2015) xxx–xxx
of the packet and there is a collision or (ii) the deferral
counter expires so that the node defers its transmission.
The probabilities of each of these situations depend on
the backoff stage as they are a function of the value of
W i and M i . Thus, we compute the probability of failure at
backoff stage i as shown in Eq. (11).
ðiÞ ðiÞ ðiÞ
pf ¼ p  pbo þ pdefer ;
ðiÞ
where pbo is the probability that the backoff expires in
ðiÞ
backoff stage i. On the other side, pdefer denotes the proba-
bility that the deferral counter reaches zero and a new
packet is overheard in the given backoff stage. This proba-
bility is also obtained from [11] and it is calculated as:
WX k1 
i M i X
 probability to defer a packet in a given backoff stage by
1 Mi þ l M
ð1  pb Þl pb i pb :
ðiÞ
pdefer ¼
W i þ 1 k¼1 l¼0 l ral counter with exponential distributions with means
The probability that the backoff expires is just the com-
plementary of the probability to defer (Eq. (13)).
ðiÞ
pbo ¼ 1  pdefer
ðiÞ
Thus, the total average number of slots in backoff is
obtained by summing the number of slots waiting at each
stage weighted by the probability of moving to that backoff
stage (see Eq. (14)). Note that the probability to move to a
given backoff stage is the probability to face a failure in the
previous ones. An extra term is considered for the last
backoff stage as the node returns to it after every failure.
Assuming a packet is retransmitted until it is successfully
transmitted the last term follows.
X
m1 Y
i1 Y
m1
ðjÞ
E½w ¼ E½w1  þ E½wi  pf þ E½wm 
i¼2 j¼1 j¼1
Finally, having the payload length (L), we obtain the
throughput as:
L note that the estimation is accurate for certain values of
S¼q
X pb , M i and W i . As will be shown in Section 6, the approx-
The previous expressions can be solved by using an
iterative numerical method. If compared to the previous
analytical model in [11], the renewal reward approach
allow us to reduce the complexity by removing two itera-
tion loops: the one involving the calculation of the state
probabilities of the Markov Chain as well as the one
regarding the total average MAC access delay computation.
ðiÞ
4.2. Precomputation of pdefer and E½wi 
The most computationally expensive operations of our
simplified analytical model are the computation of pdefer
and E½wi  (Eqs. (9) and (12), respectively). However, both
are a function of just three variables: pdefer ¼ f ðW i ; M i ; pb Þ
and E½wi  ¼ gðW i ; M i ; pb Þ. Since W i and M i are defined as
fixed parameters in the standards and pb 2 ½0; 1, we can
easily precompute the values of these two metrics for each
relevant value of W i , M i and pb . Alternatively, if W i and M i
are selected so as to improve performance depending on
the network conditions, it would be practical to do a one-
Please cite this article in press as: C. Cano, D. Malone, On efficiency and validity of previous Homeplug MAC performance analysis, Comput.
Netw. (2015), http://dx.doi.org/10.1016/j.comnet.2015.03.005
off precomputation of the values when these parameters
are adjusted.
The error of the approximated results can be reduced by
applying a linear interpolation between the two closest
precomputed values of pb . Results presented in this work
have been obtained using the precomputed values as they
have been found indistinguishable (using a step size equal
ð11Þ to 104 for pb ) from the ones obtained using Eqs. (9) and
(12).
ðiÞ
4.3. Exponential approximation to pdefer
We consider a further optional step in reducing the
complexity of the analytical model. We may calculate the
approximating both the backoff countdown and the defer-
ð12Þ
chosen to match the actual distributions. We approximate
the backoff counter and the deferral counter exponentially
distributed with rate bi ¼ 2=W i and ci ¼ pb =ðM i þ 1Þ,
ð13Þ respectively. Then, we can easily derive the probability to
defer as the probability that the deferral counter expires
first, that is:
ðiÞ ci
pdefer ¼ : ð16Þ
bi þ ci
Although the optional approximation considerably
ðiÞ
reduces the complexity of the calculation of pdefer in Eq.
(12), it only provides a rough estimate as shown in
Fig. 3(a) (CA3/2) and (b) (CA1/0) for an arbitrary value of
pb . The disagreement is more notable for high values of
M i . The reason for this result is the inability of the approx-
ðjÞ 1
pf ðmÞ
ð14Þ imation to model the cases in which the random backoff
1  pf selected at stage i is smaller than M i . In such a case, regard-
less of the value of pb , the backoff counter will always
expire first, yet the exponential approximation always
gives some probability that either expire first. However,
ð15Þ
imation does provide a good estimate of throughput and
delay for all CA3/2 configurations and when p < 0:4 for
CA1/0.
4.4. Extension to model the priority resolution scheme
When we consider stations with heterogeneous CAs
accessing the channel, the priority resolution scheme pre-
sented in Section 2 takes place. Using the signalling in PRS
the highest-access-category stations prevent lower-prior-
ity stations from attempting transmission. Focusing on
ðiÞ
the saturated case, highest priority packets are always
transmitted first except when a collision takes place. In
such a case, PRS are not present, thus, contention among
ðiÞ
all CAs occurs and the behaviour is dominated by the dif-
ferent contention access parameters at each CA. The beha-
viour of the network can therefore be divided in two
different cases: Case 1 when there has been a collision
and then, all stations try to contend for the channel, and
Case 2 when there has been a successful transmission
and therefore, only the highest-access-category stations
 C. Cano, D. Malone / Computer Networks xxx (2015) xxx–xxx 7

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
defer

defer
0.5 0.5
p

p
0.4 0.4

0.3 Exact Calculation 0.3 Exact Calculation

Exp. Approx. Exp. Approx.
0.2 M=0, W=7 0.2 M=0, W=7
M=1, W=15 M=1, W=15
0.1 M=3, W=15 0.1 M=3, W=31
M=15, W=31 M=15, W=63
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
pb pb

(a) CA3/2 (b) CA1/0

ðiÞ
Fig. 3. Comparison of pdefer computed following the exact calculation and the exponential approximation.

attempt transmission. Our key assumption to model the To compute the average slot duration from the point of
behaviour of the network is to consider the steady-state view of node j (aj ) we now have to consider that:
performance metrics of these two different cases in isola- X Y
tion and then compute the global performance using a ps;j ¼ su ð1  sv Þ;
u2½1;n;u–j v 2½1;n;v –u–j
weighted mean. Y
Note that the real behaviour of the network differs pe;j ¼ ð1  su Þ;
slightly from this assumption. When there is a collision, u2½1;n;u–j

the stations that will succeed in accessing the channel
are those with the lowest remaining backoff. However, to
model these dynamics requires extra complexity in the
analysis. Thus, we have chosen to maintain the simplicity
of our previous analysis and design a low-complexity ana-
lytic model, at a potential cost of reduced accuracy. We will
evaluate the results in Section 6 and show how the tenden-
cies are correctly captured and that the accuracy is well
suited to derive useful conclusions.
We first describe the extension of the previous analysis
to model nodes accessing the channel using heterogeneous
parameters and then, we show how to model Case 1 and
Case 2, followed by the description of how to obtain the
global performance based on the previous cases.
4.4.1. Heterogeneous CAs
Our analysis in Section 4.1 can be extended to model
nodes with heterogeneous channel access parameters
(for simplicity, we consider nodes to use the same frame
length). The service time of a target node j 2 ½1; n is
now:
X j ¼ E½wj aj þ ðnt;j  1ÞT c þ T s ;
where the average number of attempts to transmit a frame
for node j is nt;j ¼ 1=ð1  pj Þ, pj being the conditional colli-
sion probability experienced by the target node:
Y
pj ¼ 1  ð1  sk Þ
k2½1;n;k–j
with sj ¼ nt;j =ðE½wj  þ nt;j Þ. Note that assuming saturation
conditions now Ij ¼ 0.
pc;j ¼ 1  ps;j  pe;j : ð19Þ
Note that the rest of metrics can be computed for each
node j using Eqs. (9)–(14) but plugging the values W i;j and
Mi;j (the contention window and starting values of the
deferral counter at stage i of node j). Finally, we can com-
pute the throughput of a given node j as Sj ¼ L=X j .
4.4.2. Case 1: contention among all stations
After a collision, all stations compete for the channel
since priority resolution symbols are not present. Thus, to
obtain the network performance in this case we have to
consider that the total number of nodes n is the sum of
nodes contending at each of the defined CAs:
n ¼ nCA 3 þ nCA 2 þ nCA 1 þ nCA 0. Assume we label stations
using sub-index j starting at the highest CAs and then con-
tinue in descending order.
The channel access parameters at each backoff stage of
the different categories, which are shown in Table 1, are
captured by the model using W i;j and M i;j . Thus, the result-
ing throughput of a given station j using a certain CA is
obtained solving the model in Section 4.4.1 for Sj , which
ð17Þ can be done using an iterative solver algorithm. Observe
that this case also models the performance of the network
when heterogeneous CAs access the channel but no prior-
ity resolution mechanism takes place.
4.4.3. Case 2: contention among the highest CA
ð18Þ In contrast to the previous case, when a station success-
fully transmits, lower-priority stations are not allowed to
transmit as the highest-CA stations will make use of the
PRS to advertise the priority of their pending packet. To
obtain the performance in this case we have to solve the

Please cite this article in press as: C. Cano, D. Malone, On efficiency and validity of previous Homeplug MAC performance analysis, Comput.
Netw. (2015), http://dx.doi.org/10.1016/j.comnet.2015.03.005
8 C. Cano, D. Malone / Computer Networks xxx (2015) xxx–xxx
analytic model considering n ¼ nCAx , x being the highest CA
for which nCAx > 0.
Note in this case that the heterogeneous channel access
consideration is not needed and that the resulting perfor-
mance can be obtained using the model in Section 4.1 set-
ting W i ¼ W i;CAx and M i ¼ Mi;CAx , W i;CAx and M i;CAx being W i
and M i of CAx. Sj will then represent the throughput of
node j, with j 6 nCAx .
4.4.4. Global performance
To obtain the global performance, we consider the net-
work to behave as a two-state, discrete-time Markov chain,
in which the states represent Case 1 and Case 2 and transi-
tions occur after packet transmissions. Considering that a
successful transmission moves the system to Case 2 and
that collisions move the system to Case 1, the transition
probabilities are:
P Case 1!Case 2 ¼ 1  pCase 1 ; P Case 1!Case 1 ¼ pCase 1 ;
P Case 2!Case 1 ¼ pCase 2 ; PCase 2!Case 2 ¼ 1  pCase 2 ;
where pCase 1 and pCase 2 are the probabilities of collision
given that a transmission occurs in Case 2 (Eq. (21)) and
Case 2 (Eq. 22), respectively.
P Q
1 s Case 1
u2½1;n u v 2½1;n;v –u ð1  sv
Case 1
Þ mostly empty or saturated conditions. On one hand, if a
pCase 1 ¼ Q ð21Þ
1 u2½1;n ð1  s
Case 1 Þ
u
P Q
1 s Case 1
u2½1;nCAx  u v 2½1;nCAx ;v –u ð1  sv
Case 1
Þ
pCase 1 ¼ Q :
u2½1;nCAx  ð1  su
1 Case 1 Þ
Solving for the state probabilities and considering that
the time elapsed at each state can be approximated to be
similar, the resulting throughput of CAx is:
pCase2 SCase1 þ ð1  pCase1 ÞSCase2
j j
Sj6nCAx ¼ : ð23Þ
1  pCase 1 þ pCase 2
Note that the lower-access-category stations are not
effectively attempting transmission in Case 2. Thus, the
only contribution is the one from Case 1:
pCase2 SCase1
j 5.1.1. Which solution corresponds to the performance of the
Sj>nCAx ¼ : ð24Þ
1  pCase 1 þ pCase 2
Observe that to solve for the global performance we
need to solve two iterative numerical algorithms.
However, by avoiding the calculation of all state probabili-
ties this analytic model remains computationally efficient
even though it considers heterogeneous CAs.
5. On the performance results right before saturation
Before the performance evaluation, in this section, we
demonstrate that the results presented in [11] just prior
to predicted saturation correspond to a transitory phase
of the network. To do so we refer to previous literature
regarding the decoupling approximation and to show that
the coupled system of queues under infinite buffer size is
unstable in this regime. Thus, with an infinite queue size,
higher throughput than the one found in saturation, as
Please cite this article in press as: C. Cano, D. Malone, On efficiency and validity of previous Homeplug MAC performance analysis, Comput.
Netw. (2015), http://dx.doi.org/10.1016/j.comnet.2015.03.005
predicted in [11], cannot be maintained in the long term.
Furthermore, we support this demonstration by sim-
ulations in which we show, under certain circumstances,
the extremely long time during which the network
remains in the transitory period before moving to the
long-term behaviour.
5.1. The decoupling approximation
The decoupling approximation allows to make the
analysis tractable by assuming that each queue can be
modelled independently of the dynamics of the rest of
queues in the network. However, one has to be careful
when interpreting the results obtained. As already
observed in [27], unsaturated analytical models that
decouple the queue dynamics can provide two different
solutions in certain regimes. In particular, when the packet
arrival rate is slightly higher than the maximum load that
the system could serve in saturation. These analytical mod-
ð20Þ els do not consider the number of instantaneous contend-
ing stations, i.e., the number of queues that have at least a
packet buffered at the same time. Neglecting this fact
makes it impossible to identify the actual regime at which
the queues operate in this region. This is caused by the
possibility of facing two extreme cases: the queues being
reduced number of nodes are contending for the channel,
the channel access delay is small, thus, making unlikely
that a high number of packets accumulate in the buffer
for transmission. This implies higher throughput than
ð22Þ
found in saturation, as the conditional collision probability
is reduced. On the other hand, if a higher number of nodes
are contending, the channel access delay increases and so
does the probability of having more packets accumulating
for transmission. The effect of the latter case, under infinite
buffer size, is the saturation of the buffer. Analytical mod-
els that do not consider the number of instantaneous con-
tending stations are not always able to differentiate among
the two different solutions and thus, iterative algorithms
may converge to either, depending on the starting
parameters.
network in the long term?
As already conjectured in [27], when considering infi-
nite buffer size, the lowest throughput solution is the one
that corresponds to the long-term behaviour of the net-
work, while the other solution is just transitory. This has
been demonstrated in [29], where a coupled system of par-
allel queues with infinite buffer size and state-dependent
service rates is analysed. The system of queues is found
to be unstable (the stability limit is surpassed) when the
packet arrival rate (k) is higher than the service rate that
can be achieved when all nodes are simultaneously con-
tending for the channel. This service rate corresponds to
the one found in saturation. Therefore, when the packet
arrival rate is higher than the service rate in saturation,
all queues will eventually have an increasing number of
packets. Thus, the long-term performance of the network
corresponds to that in saturation (all nodes will have at
least a packet buffered for transmission). The results
 C. Cano, D. Malone / Computer Networks xxx (2015) xxx–xxx 9

presented in [11] show higher throughput than the one 60

CA3/2
found in saturation when the stability limit is slightly sur- CA1/0

Saturated Service Rate [packets/s]

passed. However, given that the queues are unstable in 50
that regime, the long-term throughput should correspond
to the one in saturation instead. Thus, as will be shown 40
in Section 6, when two solutions are obtained from the
analysis, only the lower-throughput one corresponds to 30
the long-term performance of the network, the higher-
throughput solution identified in [11] corresponds just to 20
a transitory phase of the network and not to the long-term
throughput.
10
In order to identify the stability limit regime, Fig. 4
shows, for different number of nodes contending for the
0
channel and access categories, the service rate obtained 10 15 20 25 30 35 40 45 50
in saturation conditions. The parameters used are shown Number of stations (n)
in Table 2 and correspond to those in [11], which are
defined in Homeplug 1.0 [3]. Note that the service rate in Fig. 4. Service rate in saturated conditions.
saturated conditions is just lsat ¼ 1=X sat , where X sat
denotes here the service time considering that all nodes
Table 2
always have a packet ready for transmission. From [29] Parameters Homeplug 1.0.
(for a homogeneous system), we identify the stability con-
Parameter Value in Homeplug 1.0
dition of the system of coupled queues as k < lsat . As will
be shown in Section 6, when the stability condition of Data rate (R) 14 Mbps
Frame transmission time (T fra )/L 1153.5 ls/1500 bytes
the network is not satisfied, there is a potential for obtain-
ACK transmission time (T res ) 72 ls
ing two solutions from the analytical model. The stability Slot time (r) 35.84 ls
limit is surpassed if k is higher than the service rate Data-ACK interframe space (RIFS) 26 ls
depicted. It is important to emphasise that the potential Contention interframe space (CIFS) 35.84 ls
for obtaining two solutions decreases when k  lsat as Tx. indication slots (PRS0 ¼ PRS1) 35.84 ls

the probability of having a large number of nodes are

simultaneously contending for the channel (i.e., stations
with at least a packet buffered for transmission) increases.
Thus, the initial phase in which only few stations are con-
tending and higher throughput than in saturation is
obtained may then be negligible and we find the analytical
model converges to a unique solution. This will be shown
in Section 6. The effect of this in terms of the duration of
the long transitory phase will be evaluated in the next
subsection.
When using iterative numerical solvers, setting certain
starting parameters can provide a particular solution
among the two possible ones. Setting the metrics assuming
a highly-loaded network starts the numerics closer to the
saturated solution. On the contrary, assuming lightly-
loaded conditions, starts the iterative loop closer to the
transitory phase. When there is only one solution, the sol-
ver converges to it independently of starting under any of
these initial conditions. We will use this technique in
Section 6,, to obtain both solutions from the analysis.
This technique, although useful to distinguish the two
solutions, is not able to differentiate them when both are
very close. However, we know that after surpassing the
stability limit (as described above) the throughput that
should be obtained is the one corresponding to saturation.
5.2. The long transitory phase in simulations
When k  lsat , the system rapidly moves to the long-
term behaviour. However, performing simulations right
after the stability limit can provide wrong results since,
as we have found in this work, the length of the transitory
phase can be extremely long. We have found that if the
experiments are started with the queues empty, it can take
a long time to reach the long-term performance since the
system has to reach a point at which a large number of
nodes are simultaneously contending for the channel and
start to have their queues filled with an increasing number
of packets. The possibility of a long transitory phase for
random MAC protocols in this regime was postulated in
[30], however no experimental findings or proof was pro-
vided. Here we take the step of showing by simulations
that for Homeplug MAC this is exactly the case.
For that purpose we have run long simulations with
load just above the stability limit with queues long enough
to be considered infinite (1000 packets). From the sim-
ulations we have measured the throughput and the queue
size (maximum, average and minimum) of all the nodes in
the network in 1s-long time intervals. Simulation results
are obtained using a custom simulator based on the
SENSE framework [31]. We have considered the case in
which 50 nodes with packets belonging to CA3/2 access
categories are accessing the channel, as larger numbers
of nodes help us highlight the discrepancy between the
previously presented results (see [11] Figs. 6 and 7) and
the saturated solution. We begin the simulations with
the queues empty.
Observe in Fig. 5(a) and (c) that, for different values of
the packet arrival rate, there is a sharp change from the
transitory behaviour to the long-term solution. In
Fig. 5(b) and (d), we zoom in on the queue behaviour
around this sharp change. We see that the queues start
to fill with packets until the maximum size is reached.

Please cite this article in press as: C. Cano, D. Malone, On efficiency and validity of previous Homeplug MAC performance analysis, Comput.
Netw. (2015), http://dx.doi.org/10.1016/j.comnet.2015.03.005
10 C. Cano, D. Malone / Computer Networks xxx (2015) xxx–xxx

6
x 10
1000
6
900
5.5
800
maximum→

Queue Size [packets]

5
Throughput [bits/s]

700
4.5 600
←average
4 500

3.5 400
←minimum
300
3
200
2.5
100
2
0
0 5 10 15 20 4.85 4.9 4.95 5
Time [h] Time [h]

(a) λ = 8 packets/s (b) λ = 8 packets/s

6
x 10
1000
6
900
5.5
800
maximum→
Queue Size [packets]
5
Throughput [bits/s]

700
4.5 600
←average
4 500

3.5 400
←minimum
3 300

200
2.5
100
2
0
0 5 10 15 20 25 30 15.45 15.5 15.55 15.6 15.65
Time [h] Time [h]

(c) λ = 7.5 packets/s (d) λ = 7.5 packets/s

Fig. 5. Evolution of throughput and queue size (CA3/2, n ¼ 50 nodes).

The conjecture that the transitory phase is of long duration saturated and then start taking the statistics of the perfor-
is also supported, as we observe a duration of the order of mance metrics of interest. However, we suggest starting
hours in the runs shown. The two solutions for throughput the simulations with a number of packets preloaded in
obtained from analysis (these will be shown in Section 6, the queues as a more practical way to force the system
Fig. 8) correspond to the two phases depicted in Fig. 5(a) to enter into the long-term operating state. If the queues
and (c). However, as can be seen in Fig. 5(a) and (c), the are stable, nodes will be able to release these preloaded
long-term performance corresponds to the lower-through- packets in a reasonable amount of time. If the queues are
put instead of the higher-throughput one, in contrast to the unstable, we have started the simulation closer to their
results presented in [11]. stationary regime, and will see the long-term throughput
It is important to emphasise that the results presented more quickly. We use this technique to avoid the transitory
in Fig. 5 are each from a single simulation run. The time period in the next section in order to suppress the effect of
at which the change in behaviour occurs varies in different the transitory on performance results.
runs due to its stochastic nature. To characterise the length
of the transitory period one must perform analysis con- 6. Performance evaluation
sidering the coupled dynamics of the system of queues.
Due to its difficulty, especially considering the complexity In this section, we present the validation of our analysis.
of the Homeplug MAC access procedure, we consider this The validation for the homogeneous case is divided accord-
analysis out of the scope of this work. ing to saturated and unsaturated conditions. The later also
Because of this long transitory period, care is required show the two different solutions the analytical model may
in designing the simulations. One way to obtain the long- provide as well as the long-term performance to which the
term performance is to start with the queues empty, run system converges to, obtained from simulations. We also
the simulation for a long time until the queues are evaluate the effect of the deferral counter on misprediction

Please cite this article in press as: C. Cano, D. Malone, On efficiency and validity of previous Homeplug MAC performance analysis, Comput.
Netw. (2015), http://dx.doi.org/10.1016/j.comnet.2015.03.005
 C. Cano, D. Malone / Computer Networks xxx (2015) xxx–xxx 11

errors. We then validate our analytical framework for Exact Calculation

modelling the strict priority resolution scheme and show Exp. Approx.

how starvation of lower-access category traffic can be 0.25

MAC Acccess Delay [s]

solved resorting to the different contention access parame-
ters alone. In the last subsection, we show the complexity 0.2

reduction of our simplified analysis.

Simulations are performed, as in the last section, using 0.15
the SENSE framework and parameters shown in Table 2. CA3/2
Results show averages from simulation runs of 10,000 s. 0.1

CA1/0
6.1. Saturated conditions 0.05

Throughput and channel access delay metrics in sat- 0

0 0.2 0.4 0.6 0.8 1
uration conditions for different number of nodes are p
shown in Fig. 6(a) and (b), respectively. Results from (i)
the exact computation (obtained from precomputing the Fig. 7. Channel access delay vs. p.
ðiÞ
values of pdefer and E½wi ), (ii) the exponential approx-
ðiÞ
imation to pdefer and (iii) simulations are depicted. approximation is accurate for access categories CA3/2 for
Observe that the results obtained from the exact calcula- all configurations, while for access categories CA1/0 it pro-
tion are in agreement with the performance predictions vides a good estimate when p < 0:4.
and simulations presented in [11] in Figs. 4 and 5 (note
that the only difference is our T s ¼ T c consideration). 6.2. Unsaturated conditions
Moreover, we also see how the optional exponential
approximation is able to accurately predict both through- Results in unsaturated conditions for n = 10, 30 and 50
put and channel access delay in all the cases considered and varying the packet arrival rate k are shown in Fig. 8.
for access categories CA3/2. For access categories CA1/0, Values of throughput and channel access delay for the dif-
the exponential approximation is accurate for a small ferent access categories obtained from the exact com-
number of contending nodes (less than 15). For a large putation and simulations are depicted.
number of nodes, the approximation provides a rough For the analytical results, we show two solutions,
estimate. corresponding to beginning the numerics for solving the
To study when the exponential approximation is useful, model with I ¼ 0 slots (no idle periods between a depar-
Fig. 7 shows the channel access delay vs. the conditional ture and the next packet arrival, labelled Analysis 1) or
collision probability, obtained using the exact calculation I ¼ 1000 slots (long idle periods between departures and
and the exponential approximation. This plot allows us to next packet arrivals, labelled Analysis 2). Observe in Fig. 8
identify when the exponential approximation gives a good that the model may converge to different solutions for
estimate, independent of the exact configuration, since the throughput for a given k, as described in Section 5. As
channel access delay as a function of the conditional colli- expected, just after the stability limit, the analytical model
sion probability is independent of the number of nodes provides two different solutions depending on the starting
contending [32]. Thus, we observe that the exponential values we provide to the iterative loop. We see that

6
x 10
9
Exact Calculation
Exp. Approx.
8
Simulations
0.25
7
MAC Acccess Delay [s]

CA1/0
Throughput [bits/s]

6 0.2

5
0.15
4
CA3/2 CA3/2
3 0.1

2 CA1/0
0.05
Exact Calculation
1 Exp. Approx.
Simulations
0 0
0 10 20 30 40 50 0 10 20 30 40 50

Number of stations (n) Number of stations (n)

ðiÞ
Fig. 6. Performance results in saturation conditions. Comparison among the exact calculation, the exponential approximation of pdefer and simulations.

Please cite this article in press as: C. Cano, D. Malone, On efficiency and validity of previous Homeplug MAC performance analysis, Comput.
Netw. (2015), http://dx.doi.org/10.1016/j.comnet.2015.03.005
12 C. Cano, D. Malone / Computer Networks xxx (2015) xxx–xxx

Analysis 1 provides the lowest throughput, while Analysis 2 6.3. Effect of the deferral counter
shows higher throughput.
We use here the technique of preloading the queues as We now evaluate the effect of the deferral counter on the
already discussed in the last section. Thus, we have aver- difference between the two different solutions obtained
aged the values of 10,000 s-long simulations with the from the analytical model right after the stability limit.
queues preloaded with 50 packets. Fig. 8 shows the results These results allow us to quantify the potential mispredic-
from the long-term performance obtained after the transi- tion errors. We consider (i) the results with no deferral
tory phase in simulations and it can be observed that (i.e., corresponding to a DCF-like access procedure [7]) and
agreement is found with results from Analysis 1. (ii) increasing the backoff stage every time a packet from a
Note how, excluding the results right after saturation neighbouring node is overheard (i.e., the proposal presented
(as we already discussed), the results obtained are also in in [21] for improving the performance of Homeplug MAC).
agreement with the outcomes obtained in [11] in Figs. 6–
9, validating, as in the saturated case, the accuracy of our 6.3.1. No deferral
proposed analysis. To obtain the results without the deferral counter (i.e.,
Finally, we compare the optional exponential approx- M i ¼ 1; 8i), we run the analytical model described in
imation and the exact calculation (both starting with ðiÞ
Section 4 with pdefer ¼ 0; 8i and E½wi  ¼ ðW i þ 1Þ=2. Note
I ¼ 0 slots) in unsaturated conditions in Fig. 9. When the that the performance obtained will match that of DCF with
system is saturated, we obtain the accuracy presented in the same parameters. However, for comparison purposes,
the last subsection. Here it can be observed how the we use the MAC parameters of Homeplug MAC shown in
exponential approximation is able to predict the saturation Table 2. We start the numerics for solving the model as
point with a difference, for all cases evaluated, smaller in the last subsection: Analysis 1 corresponds to an initial
than Dk ¼ 2 packets/s. value of I ¼ 0 slots and Analysis 2 corresponds to the results

6
x 10
9
Analysis 1
Analysis 2
8
0.3 Simulations

7 n=50
MAC Access Delay [s]

n=10 0.25
Throughput [bits/s]

5 0.2
n=30
4 0.15

3 n=50 n=30
0.1
2
Analysis 1 0.05
1 Analysis 2 n=10
Simulations
0 0
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80

λ [packets/s] λ [packets/s]

(a) Throughput CA3/2 (b) MAC access delay CA3/2

6
x 10
9
Analysis 1
Analysis 2
8
0.3 Simulations
n=10
7
MAC Access Delay [s]

0.25
Throughput [bits/s]

n=30
6
n=50 0.2
5

4 0.15
n=50
3
0.1
2 n=30

Analysis 1 0.05
1 Analysis 2 n=10
Simulations
0 0
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80

λ [packets/s] λ [packets/s]

(c) Throughput CA1/0 (d) MAC access delay CA1/0

Fig. 8. Performance results in unsaturated conditions. Comparison among the two solutions derived from the exact analysis and simulations.

Please cite this article in press as: C. Cano, D. Malone, On efficiency and validity of previous Homeplug MAC performance analysis, Comput.
Netw. (2015), http://dx.doi.org/10.1016/j.comnet.2015.03.005
 C. Cano, D. Malone / Computer Networks xxx (2015) xxx–xxx 13

6
x 10
9
Exact Calculation Exact Calculation
Exp. Approx. Exp. Approx.
8
0.3
7
Throughput [bits/s]

MAC Access Delay [s]

n=10 n=50
0.25
6

5 0.2
n=30
4 0.15

3 n=50
0.1 n=30
2
0.05
1 n=10

0 0
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
λ [packets/s] λ [packets/s]

(a) Throughput CA3/2 (b) MAC access delay CA3/2

6
x 10
9
Exact Calculation Exact Calculation
Exp. Approx. Exp. Approx.
8
0.3
n=10
7
Throughput [bits/s]

MAC Access Delay [s]

n=30 0.25
6

5 n=50 0.2

4 n=50
0.15

3
0.1
2 n=30

0.05
1 n=10

0 0
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
λ [packets/s] λ [packets/s]

(c) Throughput CA1/0 (d) MAC access delay CA1/0

ðiÞ
Fig. 9. Performance results in unsaturated conditions. Comparison among the exact analysis and the exponential approximation to pdefer .

starting with I ¼ 1000 slots. The two different solutions for increased number of backoff stages and larger contention
throughput obtained from the analytical model are shown windows. The improvement in the saturation throughput
in Fig. 10. Note that compared to the results depicted in by setting M i ¼ 0; 8i has already been studied for different
Fig. 8, the difference between the two solutions provided configurations of i, W and payload lengths (see Fig. 3 in
by the analytical model is higher in Fig. 10. The saturation [21]). However, in order to perform a direct comparison
throughput obtained with no deferral is smaller compared with the outcomes presented in previous sections, we
to the one obtained using the deferral counter. Thus, with show in Fig. 11 the results using the same parameters con-
no deferral (Fig. 10), the saturation throughput (Analysis 1) sidered throughout this work (i.e., those shown in Table 1
and the throughput obtained under a lower contention and 2) instead of referring to results in [21]. Moreover, this
assumption (Analysis 2) show a bigger difference than the configuration allows us to evaluate the benefit of increas-
one found with deferral (Fig. 8). Thus, under the same con- ing the backoff stage after every packet overheard when
figuration, protocols such as the DCF have the potential to the number of backoff stages are set to those recom-
be affected by higher misprediction errors than Homeplug mended by the standard.
MAC, especially when small values of W are used as in the Observe that, under these particular settings, the sat-
case of CA3/2 categories. uration throughput compared to Fig. 8 is substantially
improved but the potential misprediction errors are still
6.3.2. Deferring always after overhearing considerable.
In accordance with the conclusions drawn for no defer-
ral, we expect that deferring always after overhearing will 6.4. Effects of the priority resolution scheme
show a smaller difference among the two solutions as this
setting increases the saturation throughput with higher Results using the priority resolution scheme while vary-
contending nodes, especially when considering an ing the number of CA3 stations (nCA3 ) and setting

Please cite this article in press as: C. Cano, D. Malone, On efficiency and validity of previous Homeplug MAC performance analysis, Comput.
Netw. (2015), http://dx.doi.org/10.1016/j.comnet.2015.03.005
14 C. Cano, D. Malone / Computer Networks xxx (2015) xxx–xxx

6 6
x 10 x 10
9 9

8 8

7 7
Throughput [bits/s]

Throughput [bits/s]
n=10
6 6
n=10
5 5
n=30
4 4
n=50
3 n=30 3

2 2
n=50
1 Analysis 1
1 Analysis 1
Analysis 2 Analysis 2
0 0
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80

λ [packets/s] λ [packets/s]
(a) Throughput CA3/2 (b) Throughput CA1/0
Fig. 10. Throughput in unsaturated conditions with no deferral (Mi ¼ 1; 8i). Comparison among the two solutions derived from the exact analysis.

6 6
x 10 x 10
9 9

8 8
n=10
n=10
7 7
n=30
Throughput [bits/s]

Throughput [bits/s]

6 6 n=50
n=30
5 5

4 n=50 4

3 3

2 2

1 Analysis 1
1 Analysis 1
Analysis 2 Analysis 2
0 0
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80

λ [packets/s] λ [packets/s]

(a) Throughput CA3/2 (b) Throughput CA1/0

Fig. 11. Throughput in unsaturated conditions deferring always after overhearing (M i ¼ 0; 8i). Comparison among the two solutions derived from the exact
analysis.

nCA2 ¼ nCA1 ¼ nCA0 ¼ 1 and nCA2 ¼ nCA1 ¼ nCA0 ¼ 10 are
shown in Fig. 12 for CA3 stations and in Fig. 13 for CA2-
CA0. Results are obtained using the analysis in
Section 4.4 and simulations. Note the near starvation faced
by lower-priority stations, whose throughput is less than a
tenth of the high priority, and that how this effect is ame-
liorated at increased high-priority contention until a cer-
tain point at which contention is more severe.
Regarding the accuracy of the analysis, observe in
Figs. 12 and 13 that the absolute error is small (<300 kbps
for the higher-priority and <100 kbps for the lower-prior-
ity). Additionally, we can see that our assumption in
Section 4.4 is more accurate for increased CA3 contention
than for lower contention levels. Recall that we have
neglected the remaining backoff of stations after a colli-
sion. We believe the inaccuracy of the analysis for reduced
contention is caused by the probability of CA3 stations to
have a smaller remaining backoff, due to their higher
chance to transmit that results in smaller probability of
reaching higher backoff stages. On the contrary, as the con-
tention increases by adding a higher number of CA3 sta-
tions, their remaining backoff is also higher on average
and their probability to acquire the channel becomes more
balanced compared to lower priorities. Thus, the analysis is
more accurate in this regime as it assumes steady-state
conditions for each case considered: (1) contention among
all stations and (2) contention among the highest CA only.
Importantly, our analysis is able to closely capture the
main tendencies previously observed in [20,6] by sim-
ulations and experimental evaluation, respectively. These
are (1) high-priority traffic severely monopolises channel
resources and (2) increased but limited contention of
high-priority stations releases channel resources to lower
access categories. However, in real deployments, we have
observed that the second tendency is ameliorated by the
capture effect, see [6].

Please cite this article in press as: C. Cano, D. Malone, On efficiency and validity of previous Homeplug MAC performance analysis, Comput.
Netw. (2015), http://dx.doi.org/10.1016/j.comnet.2015.03.005
 C. Cano, D. Malone / Computer Networks xxx (2015) xxx–xxx 15

6 6
x 10 x 10

8 Analysis − CA3 8 Analysis − CA3

Simulations − CA3 Simulations − CA3
Aggregated Throughput [bits/s]

Aggregated Throughput [bits/s]

7 7

6 6

5 5

4 4

3 3

2 2

1 1

0 0
0 10 20 30 40 50 0 10 20 30 40 50
Number of CA3 stations Number of CA3 stations
(a) nCA2 = nCA1 = nCA0 = 1 (b) nCA2 = nCA1 = nCA0 = 10

Fig. 12. Results in saturated conditions for different configurations of nCAs using the priority resolution scheme (results for CA3).

5 5
x 10 x 10
2.5 6
Analysis − CA2 Analysis − CA2
Analysis − CA1/CA0 Analysis − CA1/CA0
Aggregated Throughput [bits/s]

Simulations − CA2 Simulations − CA2

Aggregated Throughput [bits/s]

5
2 Simulations − CA1/CA0 Simulations − CA1/CA0

4
1.5

1
2

0.5
1

0 0
0 10 20 30 40 50 0 10 20 30 40 50
Number of CA3 stations Number of CA3 stations
(a) nCA2 = nCA1 = nCA0 = 1 (b) nCA2 = nCA1 = nCA0 = 10

Fig. 13. Results in saturated conditions for different configurations of nCAs using the priority resolution scheme (results for CA2 and CA1/0).

6.5. Differentiation through contention access parameters have also tested three different configurations of M i and
alone W i . Fig. 14(a) shows the results for the configuration
suggested in the standards (shown in Table 1). In
The performance obtained using the priority resolution Fig. 14(b) a lower M i configuration is used for CA1/0 in
scheme and shown in Figs. 12 and 13 may have an impact which M 1 ¼ 0 and M i ¼ 1; 2 6 i 6 4. In contrast, a higher
on user satisfaction as lower-priority traffic (including CA2, W i configuration for CA1/0 is used in Fig. 14(c), values of
which is used to transmit video) faces starvation due to the the contention window are set to W 1 ¼ 15; W 2 ¼ 31;
strict priority resolution scheme. We evaluate here how W 3 ¼ 63; W 4 ¼ 127. The per-station throughput using the
traffic differentiation among different priorities via the priority resolution scheme is also shown as reference and
contention access parameters alone, that is, not imple- labelled as Ref Analysis PRS (note that these results corre-
menting the priority resolution scheme, can solve this spond to the ones shown in Figs. 12 and 13 but shown
problem. Note that the analytical framework in per-station here to evaluate traffic differentiation more
Section 4.4.2 can be used to model this particular case. clearly).
This can be seen as the traffic differentiation provided in First, observe for all settings in Fig. 14 that the strong
IEEE 802.11e [33] but with the extra flexibility offered by starvation when nCA3 6 5 (for which the collision probabil-
the Mi parameter. ity is small and then scarce resources are released to
Fig. 14 shows the per-station throughput for different lower-access priorities) is now solved. Results show that
configurations of nCA3 and setting nCA2 ¼ nCA1 ¼ nCA0 ¼ 10 lower-access categories are now able to effectively trans-
when the priority resolution scheme is not applied. We mit even when the number of nCA3 stations is small and

Please cite this article in press as: C. Cano, D. Malone, On efficiency and validity of previous Homeplug MAC performance analysis, Comput.
Netw. (2015), http://dx.doi.org/10.1016/j.comnet.2015.03.005
16 C. Cano, D. Malone / Computer Networks xxx (2015) xxx–xxx

5
x 10 5
4 x 10
Ref Analysis PRS − CA3 4
Ref Analysis PRS − CA3
Ref Analysis PRS − CA2
3.5 Ref Analysis PRS − CA2
Per−Station Throughput [bits/s]
3.5

Per−Station Throughput [bits/s]

Ref Analysis PRS − CA1/0
Ref Analysis PRS − CA1/0
Analysis No PRS − CA3/2
3 Analysis No PRS − CA3/2
Analysis No PRS − CA1/CA0 3 Analysis No PRS − CA1/CA0
Simulations No PRS − CA3/2 Simulations No PRS − CA3/2
2.5 Simulations No PRS − CA1/0 2.5 Simulations No PRS − CA1/0

2 2

1.5 1.5

1 1

0.5 0.5

0 0
0 10 20 30 40 50 0 10 20 30 40 50
Number of CA3 stations Number of CA3 stations
(a) Standard parameters (b) Lower Mi configuration CA1/0

5
x 10
4
Ref Analysis PRS − CA3
Ref Analysis PRS − CA2
3.5
Per−Station Throughput [bits/s]

Ref Analysis PRS − CA1/0

Analysis No PRS − CA3/2
3 Analysis No PRS − CA1/CA0
Simulations No PRS − CA3/2
2.5 Simulations No PRS − CA1/0

1.5

0.5

0
0 10 20 30 40 50
Number of CA3 stations
(c) Higher Wi configuration CA1/0
Fig. 14. Results in saturated conditions for different configurations of nCA3 and setting nCA2 ¼ nCA1 ¼ nCA0 ¼ 10 differentiating through contention access
parameters alone.

that the channel resources released to CA1–0 can be 6.6. Reduction of complexity
adjusted varying the contention parameters. Note that
the traffic differentiation provided by using the standard Here, we evaluate the reduction in complexity obtained
parameters (which are set considering that the priority using our reformulated analysis (with precomputed pdefer
ðiÞ

resolution scheme takes place and thus, are not optimised and E½wi ) and also our optional exponential approximation
to provide a high traffic differentiation by themselves) can in comparison to the analysis presented in [11]. Moreover,
be improved, especially by increasing the W i configuration we also compare the time needed to run the analytical
of lower-access categories, as shown in Fig. 14(c). In this models and 10,000 s-long simulations, as a crucial value
particular case, since we are considering saturation, no of analysis is as a fast way to predict performance metrics.
substantial gain is obtained by lowering the values of the Since the analytical models are solved by means of itera-
M i of lower-access categories alone (Fig. 14(b)). The reason tive numerical methods, a regular complexity analysis
for this is the high probabilities to be in higher backoff based on the number of operations is not useful as the
states due to high contention. number of iterations needed may vary and cannot be pre-
These results demonstrate that the strong starvation of dicted in advance. For this reason, we have obtained the
lower-access categories due to the strict priority resolution elapsed times to run the analytical models (using tic and
scheme can be solved by resorting to the traffic differentia- toc commands in Matlab R2011b) and simulations (using
tion provided by the contention access parameters alone the time command in Ubuntu). Tests for all analytical
and that various levels of differentiation can be obtained frameworks and for simulations have been performed in
by fine tuning W i and M i values depending on the traffic an Intel Core i7 at 2.8 GHz with 8 GB of RAM running a
conditions and requirements. 12.04 64-bit Ubuntu operating system. Results of a single

Please cite this article in press as: C. Cano, D. Malone, On efficiency and validity of previous Homeplug MAC performance analysis, Comput.
Netw. (2015), http://dx.doi.org/10.1016/j.comnet.2015.03.005
 C. Cano, D. Malone / Computer Networks xxx (2015) xxx–xxx
Table 3
Elapsed runtime comparison.
n Analysis in Exact Exp. approx. Simulations
[11] (s) analysis (s) (s) (s)
10 421.0 2.9 1.4 179.3
20 365.0 2.6 1.2 358.5
30 400.0 2.9 1.6 666.6
40 282.7 3.5 2.1 849.4
50 344.6 3.6 3.2 1049.9
run considering 10, 20, 30, 40 and 50 nodes using access
categories CA1/0 and in saturated conditions are depicted
in Table 3. Obviously, these results may be affected by
implementation issues. However, as can be observed, the
improvement of our simplified analysis (both the exact
analysis and the exponential approximation) is 2 orders
of magnitude. Furthermore, the analysis in [11] has been
shown to be of a similar computational expense to sim-
ulations, placing reasonable doubts on its efficiency.
Note that the long runtimes of the analytical model in
[11] are not suitable for online use, that is, using the ana-
lytical model to adapt the network parameters when
changes in a network that is up and running occur. In con-
trast, simplified analysis with reduced runtime can be
implemented in practice as a way to tune the network con-
figuration to react to new conditions, such as nodes enter-
ing/leaving the network, traffic load variations or changes
in the number of stations of a given access category.
Additionally, optimisation analysis for which several
instances of the model must be executed in order to derive
conclusions about performance for a range of conditions
and configurations are also benefited from a simplified
method of solution as the one we have presented in this
work.
7. Final remarks
In this work we have reformulated the analytical model
of the Homeplug MAC procedure presented in [11] while
maintaining its accuracy. To that end, we have applied a
renewal reward approach to model the channel attempt
rate instead of solving for the state probabilities associated
with the Markov chain. Furthermore, following our
approach, the most computationally expensive operations
can be precomputed. We have considerably reduced the
complexity of the analysis, resulting in a two orders-of-
magnitude improvement in runtime. An optional exponen-
tial approximation has also been proposed and shown to be
generally accurate for access categories CA3/2 and accurate
under a small-conditional-collision-probability condition
for access categories CA1/0.
We have used our simplified framework to model the
case of heterogeneous access categories as well as the
strict priority resolution scheme defined in Homeplug
MAC and IEEE 1901. Our results are in agreement with
simulations and experimental evaluation in a real off-
the-shelf testbed. Of particular interest is that the model
is able to capture the strong starvation faced by lower-
priority stations that may have a severe impact on user
satisfaction and that may prevent further adoption of the
Please cite this article in press as: C. Cano, D. Malone, On efficiency and validity of previous Homeplug MAC performance analysis, Comput.
Netw. (2015), http://dx.doi.org/10.1016/j.comnet.2015.03.005
17
technology. We have also shown that by resorting to the
traffic differentiation provided by using different channel
access parameters can solve this starvation problem. Our
analysis can be used by manufacturers in order to address
this issue as well as for evaluating performance of different
configurations.
Building upon previous literature on mean field analysis
of random access MAC protocols and on stability analysis
of systems of parallel queues with state-dependent service
rates, we have demonstrated that the results right before
the predicted saturation point in [11] correspond to a
higher-throughput transitory phase of the network. We
have also supported this finding by simulations and shown
the extremely long duration of the transitory period.
The reformulated analytical model has been validated
in saturated and unsaturated conditions using simulations.
Furthermore, we have found agreement (except right after
the stability limit, as previously discussed) with the results
presented in [11], establishing the good accuracy of our
proposal. In unsaturated conditions, we have shown the
two solutions that the analytical model provides right after
the stability limit and the long-term behaviour the net-
work converges to by means of simulations. As expected,
the long-term performance is in agreement with one of
the solutions provided by the analytical model.
In this article we have presented the first long-term
performance evaluation of Homeplug MAC under infinite
(or large enough to be considered infinite) buffer size con-
sidering also different access categories contending for the
channel. We have also highlighted the complexity of ana-
lysing a network right after the stability limit as analytical
models that consider the decoupling approximation can
provide two different solutions and simulations have to
be run for a long time to allow the system to move to its
long-term behaviour. Moreover, we have also provided
techniques to obtain the long-term performance, both ana-
lytical and experimentally. Computing the maximum ser-
vice rate the network can support allows us to identify
the range of packet arrival rates for which a potential for
two solutions exist. We have also suggested how to set
the starting parameters of iterative numerical solvers to
obtain both solutions. Using simulations, we have shown
that the transitory phase can be drastically ameliorated if
stations begin with a number of packets preloaded in their
buffers. Indeed, this is a useful technique to identify MAC
protocols which may be subject to a transitory phase via
simulation.
The potential occurrence of the extremely long transi-
tory phase in generic random access protocols, as well as
the characterisation of its duration, are important aspects
to be analysed. They can provide further insights into the
behaviour of networks based on random access protocols
and serve as practical information to perform experimental
evaluations. However, this study has been considered out
of the scope of this work due to the complexity of analys-
ing a system of coupled queues with state-dependent ser-
vice rates, the reader is referred to [34] instead.
We have also evaluated the effect of the deferral coun-
ter on the magnitude of misprediction errors and shown
that, under the same conditions, protocols such as the
DCF have the potential to be affected by higher
18 C. Cano, D. Malone / Computer Networks xxx (2015) xxx–xxx
misprediction errors than the Homeplug MAC. On the
other side, improvements of Homeplug MAC that increase
the saturation throughput as the number of nodes
increases show smaller misprediction errors.
Acknowledgment
This publication has emanated from research supported
in part by a research grant from Science Foundation Ireland
(SFI) under Grant No. 13/RC/2077.
References
[1] Qualcomm Powerline Solutions <http://www.qca.qualcomm.com/
products/qualcomm-powerline>.
[2] H.C. Ferreira, L. Lampe, J. Newbury, T.G. Swart, Power Line
Communications, John Wiley & Sons Ltd, 2010.
[3] Homeplug Powerline Alliance, Homeplug 1.0 Specification.
[4] IEEE Std 1901, Standard for Broadband over Power Line Networks:
Medium Access Control and Physical Layer Specifications, ANSI/IEEE
Std 1901.
[5] C. Vlachou, J. Herzen, P. Thiran, Fairness of MAC protocols: IEEE 1901
vs. 802.11, in: Proceedings of the IEEE International Symposium on
Power Line Communications and its Applications (ISPLC), 2013.
[6] C. Cano, D. Malone, When priority resolution goes way too far: an
experimental evaluation in PLC networks, in: IEEE International
Conference on Communications (ICC), 2015, accepted for publication.
[7] IEEE Std 802.11, Wireless LAN Medium Access Control (MAC) and
Physical Layer (PHY) Specifications, ANSI/IEEE Std 802.11.
[8] P.B. Velloso, M. Elias, M. Campista, D.D.O. Cunha, L.H.M.K. Costa,
O.C.M.B. Duarte, Analyzing the Performance of Wireless Local Area
Networks With An Improved Collision Avoidance Mechanism,
Technical Report.
[9] L. Jian, S. Qiang, The study on the performance of backoff algorithms
in multihop power line communication networks, in: Proceedings of
the Third International Conference on Measuring Technology and
Mechatronics Automation (ICMTMA), vol. 3, 2011, pp. 974–977.
[10] C. Cano, D. Malone, Evaluation of the backoff procedure of Homeplug
MAC vs. DCF, in: Proceedings of the IEEE International Symposium
on Personal, Indoor and Mobile Radio Communications (PIMRC
2013), 2013.
[11] M.Y. Chung, M.-H. Jung, T.-J. Lee, Y. Lee, Performance analysis of
HomePlug 1.0 MAC with CSMA/CA, IEEE J. Select. Areas Commun. 24
(7) (2006) 1411–1420.
[12] S.-G. Yoon, J. Yun, S. Bank, Adaptive contention window mechanism
for enhancing throughput in HomePlug AV networks, in:
Proceedings of the 5th IEEE Consumer Communications and
Networking Conference (CCNC), 2008, pp. 190–194.
[13] M. Koutny, P. Mlynek, P. Mrakava, Homeplug simulation model for
analysis of the rounding functions, in: Proceedings of the 34th
International Conference on Telecommunications and Signal
Processing (TSP), 2011, pp. 218–221.
[14] P.J. Piñero-Escuer, J. Malgosa-Sanahuja, P. Manzanares-Lopez, J.P.
Muñoz-Gea, Homeplug-AV CSMA/CA evaluation in a real in-building
scenario, IEEE Commun. Lett. 15 (6) (2011) 683–685.
[15] M. Koutny, P. Mlynek, J. Misurec, Analysis of CSMA/CA used in power
line communication, in: Proceedings of the 36th International
Conference on Telecommunications and Signal Processing (TSP),
2013, pp. 385–388.
[16] P.J. Pi nero, J.A. Cortés, F.J. Canete, L. Diez, P. Manzanares, J. Malgosa,
A realistic HomePlug-AV simulator for in-home network services
planning, in: Proceedings of the Global Telecommunications
Conference (GLOBECOM 2011), 2011, pp. 1–5.
[17] S.-G. Yoon, D. Kang, S. Bahk, Multichannel CSMA/CA protocol for
OFDMA-based broadband power-line communications, IEEE Trans.
Power Deliv. 28 (2013) 2491–2499.
[18] A. Kumar, E. Altman, D. Miorandi, M. Goyal, New insights from a
fixed point analysis of single cell IEEE 802.11 WLANs, in:
Proceedings of the IEEE International Conference on Computer
Communications (INFOCOM), vol. 3, 2005, pp. 1550–1561.
[19] G. Bianchi, I. Tinnirello, Remarks on IEEE 802.11 DCF performance
analysis, IEEE Commun. Lett. 9 (8) (2005) 765–767.
[20] C. Cano, D. Malone, Performance evaluation of the priority resolution
scheme in PLC networks, in: Proceedings of the IEEE International
Symposium on Power Line Communications and its Applications
(ISPLC), 2014.
Please cite this article in press as: C. Cano, D. Malone, On efficiency and validity of previous Homeplug MAC performance analysis, Comput.
Netw. (2015), http://dx.doi.org/10.1016/j.comnet.2015.03.005
[21] M.E.M. Campista, L.H.M.K. Costa, O.C.M.B. Duarte, Improving the
data transmission throughput over the home electrical wiring, in:
30th Anniversary on the Proceedings of the IEEE Conference on Local
Computer Networks, 2005, 2005, pp. 318–327.
[22] C. Vlachou, A. Banchs, J. Herzen, P. Thiran, Performance analysis of
MAC for power-line communications, in: Proceedings of the ACM
SIGMETRICS Conference, 2014.
[23] C. Vlachou, A. Banchs, J. Herzen, P. Thiran, On the MAC for Power-
Line Communications: Modeling Assumptions and Performance
Tradeoffs, in: Technical Report, 2014.
[24] E. Kriminger, H. Latchman, Markov chain model of homeplug CSMA
MAC for determining optimal fixed contention window size, in:
Proceedings of the 2011 IEEE International Symposium on Power
Line Communications and its Applications (ISPLC), 2011, pp. 399–
404.
[25] K. Medepalli, F.A. Tobagi, System centric and user centric queueing
models for IEEE 802.11 based wireless LANs, in: Proceedings of the
2nd International Conference on Broadband Networks, 2005,
BroadNets 2005, 2005, pp. 612–621.
[26] B. Bellalta, M. Oliver, M. Meo, M. Guerrero, A simple model of the
IEEE 802.11 MAC protocol with heterogeneous traffic flows, in:
Proceedings of the IEEE Eurocon, 2005.
[27] K.R. Duffy, Mean field Markov models of wireless local area
networks, Markov Process. Related Fields 16 (2) (2010) 295–328.
[28] K.D. Huang, K.R. Duffy, D. Malone, On the validity of IEEE 802.11
MAC modeling hypotheses, IEEE/ACM Trans. Network. 18 (6) (2010)
1935–1948.
[29] S. Borst, M. Jonckheere, L. Leskelä, Stability of parallel queueing
systems with coupled service rates, Discrete Event Dyn. Syst. 18 (4)
(2008) 447–472.
[30] K.H. Suleiman, T. Javidi, M. Liu, S. Kittipiyakul, The Impact of MAC
Buffer Size on the Throughput Performance of IEEE 802.11, Technical
Report.
[31] G. Chen, J. Branch, M. Pflug, L. Zhu, B. Szymanski, SENSE: a sensor
network simulator, Adv. Pervas. Comput. Network. (2004) 249–267.
[32] D. Malone, K.R. Duffy, D. Leith, Modeling the 802.11 distributed
coordination function in nonsaturated heterogeneous conditions,
IEEE/ACM Trans. Network. 15 (1) (2007) 159–172.
[33] IEEE Std 802.11e, Wireless LAN Medium Access Control (MAC) and
Physical Layer (PHY) Specifications, Amendement: Medium Access
Control (MAC) for Quality of Service Enhancements, ANSI/IEEE Std
802.11e.
[34] C. Cano, D. Malone, Modeling, Analysis and Impact of a Long
Transitory Phase in Random Access Protocols abs/1405.5753 <http://
arxiv.org/abs/1405.5753>.
Cristina Cano obtained the Telecommunica-
tions Engineering Degree at the Universitat
Politecnica de Catalunya (UPC) in February
2006. Then, she received her M.Sc. (2007) and
Ph.D. (2011) on Information, Communication
and Audiovisual Media Technologies from the
Universitat Pompeu Fabra (UPF). Since July
2012, she has been working as a research
fellow at the Hamilton Institute (NUIM) in
wireless networks, sensor networks, power
line communications and MAC layer design.
David Malone received B.A. (mod), M.Sc. and
Ph.D. degrees in mathematics from Trinity
College Dublin. During his time as a post-
graduate, he became a member of the FreeBSD
development team. He is currently a SFI
Stokes Lecturer at the Hamilton Institute, NUI
Maynooth. His interests include mathematics
of networks, network measurement, IPv6 and
systems administration. He is a coauthor of
O’Reilly’s IPv6 Network Administration.