You are on page 1of 8

DECENTRALIZED USER INFORMATION MANAGEMENT SYSTEMS

FOR PEER-TO-PEER COMMUNICATION NETWORKS:


AN APPROACH BY NONSTATIONARY PEER-POPULATION
PROCESS
Kazuhiko Kuraya, Hiroyuki Masuyama, Shoji Kasahara, and Yutaka Takahashi
Graduate School of Informatics, Kyoto University, Japan
{kuraya, masuyama, kasahara, takahashi}@sys.i.kyoto-u.ac.jp

ABSTRACT
Recently, Skype has been receiving considerable attention as a peer-to-peer (P2P)
Internet telephony. In Skype, a voice connection is established via super nodes
chosen from among ordinary end-user nodes. User information such as an IP
address and port number of an on-line Skype node is managed by super nodes in a
decentralized manner where the number of super nodes changes dynamically
according to the number of online Skype nodes. In this paper, we analyze the
performance of this decentralized management system of user information. In our
analytical model, new nodes join the system according to a nonstationary Poisson
process, and the online-node process associated with the number of super nodes is
analyzed with a nonstationary Markov chain. We derive the system of difference-
differential equations for the probability distribution of the number of online nodes
to compute performance measures using the stationary peakedness approximation
method. Numerical examples show that the user-information management system
based on P2P can keep the quality of service (QoS) more stable than a client-server
system with a high-performance centralized server.

Keywords: decentralized system, nonstationary Markov chain, P2P, QoS, VoIP.

1 INTRODUCTION ordinary nodes when joining the network. Except


the login process, Skype services are provided over
Currently existing Internet telephony service is a hierarchical P2P network consisting of super
based on a client-server architecture with the nodes and ordinary ones. Super nodes are chosen
Session Initiation Protocol (SIP) [12]. In order to from among ordinary nodes to manage online-node
establish a voice connection between a pair of information such as IP address and port number in a
nodes, SIP employs three types of servers: registrar, decentralized manner. When an ordinary node
proxy and redirect servers. A main role of those issues a voice connection request, its lookup
servers is the IP address resolution for a call process is handled by the associated super nodes.
establishment. After establishing the voice The number of super nodes changes dynamically
connection, voice data is exchanged directly according to the number of online Skype nodes, and
between the caller and callee’s end hosts. Although this makes the Skype network more scalable and
SIP is known as a lightweight protocol, it is less more reliable than SIP-based VoIP networks.
scalable and less reliable with the increase in the In this paper, we model and study the
number of end users due to the client-server performance of this decentralized management
architecture. system of user information. Skype statistics [15]
Recently, Skype has been receiving reported that as of January, 2008, the number of
considerable attention as a peer-to-peer (P2P) VoIP online nodes ranges from 6 million to 14 million for
application. There are more than 370 million users a day. In order to take into account this online-node
and there are at most approximately 14 million dynamics, we assume that new nodes join the
online users [14], [15]. The services offered by system according to a nonstationary Poisson
Skype are VoIP, instant messaging (IM), and file- process. Then we analyze the online-node process
transfer [6]. In this paper, we focus on VoIP service associated with the number of super nodes with a
of Skype. nonstationary Markov chain. We derive the system
The Skype network consists of three of difference-differential equations for the time-
components: Skype login server, ordinary node dependent probability distribution of the number of
(Skype client), and super node. The Skype login online nodes to compute performance measures
server is a centralized server which is accessed by using the stationary peakedness approximation

1
Ubiquitous Computing and Communication Journal
method [11]. We consider the call setup time and Xie and Yang [13] investigated the impacts of
churn rate of super nodes as a measure of quality of the access capacity and routing policy of
service (QoS), and evaluate the effect of the time- autonomous systems (ASes) on the QoS of Skype.
dependent node-join process on the call setup time. They collected the statistics of Skype super nodes
We also compare the performance of the and then measured the latency and loss between
decentralized management system with that of a ordinary nodes and super nodes, quantifying the
client-server-based system. Skype QoS by MOS.
This paper is organized as follows. Section 2 Jun et al. [7] proposed two algorithms to
summarizes the related work for Skype improve Skype QoS with analytical approach. They
performance and nonstationary queueing analysis. focused on the user-search delay influenced with
In section 3, we describe the analytical model of the number of hops in a super-node overlay
user information management in Skype, analyzing a network. They leveraged graph-theoretical results
nonstationary Markov chain. We show some for the geographical distance and relationship
numerical examples in section 4. Finally, section 5 proximity of small-world networks. However, there
concludes the paper. has been little work on the fundamental analysis of
VoIP QoS over P2P-based networks.
2 RELATED WORK In general, queueing systems with
nonstationary arrival processes are difficult to
Recently, Skype has attracted much interest not analyze. Therefore, a number of approximation
only in industries but also in research communities. methods have been proposed and studied. The
Because Skype is not an open protocol, recent review on approximation methods for nonstationary
research focuses on revealing how Skype services loss queues was provided in [1]. The effect of
are provided to millions of Skype clients. Baset and nonstationary arrivals on the performance of
Schulzrinne [2] reported how Skype works in the multiserver queueing systems was numerically
Skype network consisting of the login server, studied in [5].
ordinary hosts and super nodes. From the analysis Among these approximation methods is the
of Skype network traffic, they focused on the stationary peakedness (PK) approximation [11],
following key functions: login process, user search, where the time axis is divided into subintervals and
call establishment, media transfer, codecs, and the arrival process is assumed to be stationary for
video conference. each subinterval. In other words, the arrival rate at
Guha et al. [6] investigated the characteristics each subinterval is regarded as a constant equal to
of VoIP traffic in Skype by a measurement-based the time average of the arrival rate during the
analysis, comparing the traffic characteristics in subinterval. With this value, performance measures
P2P file-sharing networks with that in traditional can be calculated by analyzing a queueing system
voice-communication networks. It was reported that with a stationary arrival process with the same
Skype users are likely to run the Skype client arrival rate.
during office hours, causing the traffic dynamics
significantly different from that observed in P2P 3 ANALYSIS OF P2P-BASED USER
file-sharing service networks. However, regardless MANAGEMENT SYSTEM
of these studies, network operation of each super
node has not been sufficiently revealed. In this section, ordinary and super nodes
The QoS of Skype was quantitatively analyzed operations in Skype are briefly summarized, and
by Chen et al. [3]. They proposed the User then the online-node process is analyzed with a
Satisfaction Index (USI), a perceptual index for nonstationary Markov chain.
quantifying Skype user satisfaction. USI is based on
the Cox proportional hazard model, defined with 3.1 Node Operation in Skype
the bitrate, jitter and Round Trip Times (RTTs) Recent experimental studies [2], [6] reported
which are obtained from a two-level sampling how Skype application works through ordinary and
approach. It was reported that user satisfaction is super nodes. In the following, we briefly
significantly affected by sender transmission speed summarize Skype’s operation for call setup and
and the superposed effect of the delay, jitter, and super node promotion. The readers are referred to
packet loss. [2], [6] for details.
Lisha and Junzhou [8] compared QoS of Skype Consider a VoIP connection establishment
with that of MSN Messenger through a between a caller and a callee over Skype. The
measurement-based analysis with respect to the ordinary node of the caller issues a VoIP
refresh rate of software, connection setup delay, connection request to his associated super node.
one-way conversation delay and Mean Opinion Skype adopts Global Index technology to search the
Score (MOS). They claimed that QoS of Skype is IP address and port number of the callee’s ordinary
well provided but that it is almost the same as QoS node. If the callee’s node information is retrieved
of MSN Messenger. and if the VoIP request is accepted, a direct VoIP

2
Ubiquitous Computing and Communication Journal
connection is established between the two ordinary 1 t[ k ]
[ k −1 ] ∫ t [ k − 1 ]
nodes. λ[ k ] = [k ]
λ ( t ) dt ,
t −t
Skype is not an open protocol, and it is not where t[k]’s are chosen in such a way that λ[k] > 0 for
clear how Global Index works for super node all k = 1, 2, ....
promotion. In [6], the authors reported that the Let π0[0] = 1 and πn [0] = 0 for all n = 1, 2, .... Let
number of super nodes varies depending on the πn [k]
(k = 1, 2,...; n = 0, 1,...) denote an
number of online Skype nodes. They also inferred
approximation to πn(t) for t ∈ (t[k−1], t[k]], i.e.,
from experiments that Skype super nodes seem to
be chosen from the ordinary nodes with much spare
bandwidth, which are publicly reachable. (π0(t), π1(t), . . . , ) ≈ (π0[k] , π1[k] , . . .),
t ∈ (t[k−1], t[k]].
3.2 Analytical Model
In the following, the event that an offline According to the PK method, we determine πn[k]’s
ordinary node becomes online is called an ordinary- (k = 1, 2,...; n = 0, 1,...) as the solution of
node arrival. In addition, we define the sojourn time
of an ordinary/super node as the period while the [k ] [k ]
0 = − λ[ k ] π 0 + μ1π1 ,   (4)
node is online.
[k ] [k ] [k ] [k ] [k ]
Noting that the number of online nodes changes 0=λ π − (λ n −1 + μ n )π n + μ n +1π n +1 ,
in a wide range during a day, we assume that n = 1,2,…, (5)
arrivals of ordinary nodes to Skype network follow
a nonstationary Poisson process with rate λ(t) (t ≥ 0). which are obtained from (2) and (3) by replacing
The sojourn times of ordinary nodes are i.i.d. the left hand sides with zeros, and πn(t)’s with
according to an exponential distribution with mean πn[k]’s. To solve (4) and (5), we assume that there
1/μ, while the sojourn times of super nodes are exists for any k = 1, 2,..., some integer n*(k) such
exponentially distributed with mean 1/μ(S). that for all n ≥ n*(k),
Because of the lack of details for super node
operation, we simply assume that each super node ρn
[k ]
= λn
[k ]
μ n < 1.
manages at most α (α ≥ 1) nodes, including the
super node itself. Thus the total capacity of the
It thus follows from (4) and (5) that for any k = 1,
network is equal to αN(S)(t), where N(S)(t) denotes
2,…,
the number of super nodes at time t. For any time t,
we assume that the number N(S)(t) of super nodes is n
minimized to be ∏ρ
[k ] [k ] [k ]
πn = π0 i , n = 1,2,K, (6)
i =1

N ( S ) (t ) = ⎡N (t ) / α ⎤. (1) ⎡ ∞ n
[k ] ⎤
−1

= ⎢1 + ∑∏ ρi ⎥ .
[k ]
π0 (7)
⎣ n =1 i =1 ⎦
3.3 Analysis
3.4 Performance Measures
Note that the process {N(t); t ≥ 0} is a
We consider the call setup time as a QoS
nonstationary birth-and-death process. Let πn(t)
measure. Let νS denote the call setup processing rate
(t ≥ 0; n = 0, 1,...) denote Pr[N(t) = n]. It then
of each super node. For simplicity, we assume that
follows from the above assumptions that
the call setup processing rate for an ordinary node
is given by N(S)(t)νS/N(t) when there are N(t) user
d
π 0 (t ) = − λ(t )π 0 (t ) + μ1π1 (t ),   (2) nodes and N(S)(t) super nodes at time t. We define
dt the call setup time TS(t) as
d
π n (t ) = λ(t )π n −1 (t ) − ( λ(t ) + μ n )π n (t ) + μ n +1π n +1 (t ),
dt N (t )
TS (t ) = . (8)
n = 1,2,…, (3) (S )
N (t )ν S

where μn = (n − ⎡n / α ⎤) μ + ⎡n / α ⎤ μ ( S ) . Since N(0) =


It then follows from (1) and (8) that for t ∈
0, π0(0) = 1 and πn(0) = 0 for all n = 1, 2,....
It is a hard task to solve numerically the (t[k−1], t[k]] (k = 1, 2, . . .), we obtain the mean call
differential-difference equations (2) and (3). setup time
Therefore we adopt the PK method [11] for the
computation of πn(t)’s. For this purpose, we 1 ⎡ N (t ) ⎤
E[TS (t )] = E⎢ ⎥
consider a sequence of times 0 = t[0] < t[1] < t[2] < .... ν S ⎣ ⎡N (t ) / α ⎤ ⎦
We then define λ[k] (k = 1, 2, . . .) as
1 ∞ n 1 ∞
n
= ∑ ∑ ⎡n / α ⎤ × π
[k ]
× π n (t ) ≈ n ,
ν S n=1 ⎡n / α ⎤ νCS n =1

3
Ubiquitous Computing and Communication Journal
450
Approximation (1/mu(S) = 20)
where πn[k]’s are computed by (6) and (7). 400 Approximation (1/mu(S) = 2)

It is well known that the dynamics of peer 350


Approximation (1/mu(S) = 0.2)
Approximation (1/mu(S) = 0.02)
joining and leaving, called churn, degrades the Simulation (1/mu(S) = 20)

Number of User Nodes


300 Simulation (1/mu(S) = 2)
performance of P2P-based network systems. Simulation (1/mu(S) = 0.2)

Therefore, the quantitative characterization of churn 250 Simulation (1/mu(S) = 0.02)

is indispensable for designing P2P-based systems. 200

In our network model, when a super node leaves the 150


P2P network, user nodes administrated by the super 100
node need to find an alternative super node. Note
that the number of super nodes leaving the network
50

per unit time is equal to the number of super nodes 0


0 4 8 12 16 20 24
arriving at the network per unit time. Focusing on Time (hours)
the super nodes leaving the network, we define the
churn rate of super nodes CS(t) as Figure 1: The number of user nodes vs. time. (α =
100, 1/μ = 2 [hours])
C S (t ) = μ ( S ) N ( S ) (t ). 4500
Approximation (1/mu(S) = 20)
4000 Approximation (1/mu(S) = 2)
It then follows from (1) and (8) that Approximation (1/mu(S) = 0.2)
3500 Approximation (1/mu(S) = 0.02)
Simulation (1/mu(S) = 20)

Number of User Nodes


⎡⎡ N (t ) ⎤ ⎤ 3000 Simulation (1/mu(S) = 2)

E[C S (t )] = μ ( S ) E ⎢⎢ ⎥⎥
Simulation (1/mu(S) = 0.2)
Simulation (1/mu(S) = 0.02)
⎣⎢ α ⎥ ⎦
2500

2000
∞ ∞
⎡n⎤ ⎡n⎤
= μ ( S ) ∑ ⎢ ⎥ × π n (t ) ≈ μ ( S ) ∑ ⎢ ⎥ × π n ,
[k ]
1500
α
n =1 ⎢ ⎥ α
n =1 ⎢ ⎥
1000

[k]
where πn ’s are computed by (6) and (7). 500

0
0 4 8 12 16 20 24
4 NUMERICAL EXAMPLES
Time (hours)

In the following, we set α = 100, that is, the Figure 2: The number of user nodes vs. time. (α =
maximum number of ordinary nodes a super node 100, 1/μ = 2 [hours])
can manage is 100. We consider time-dependent
arrivals of ordinary nodes with λ(t) = λ̄ + with rate λmax. Then, tpre is accepted as a “real”
Acos(2πt/T ), where T is the cycle length and A(> 0) arrival time with probability λ(tpre)/λmax. After that,
is the amplitude of the sinusoidal arrival process [5]. the next preliminary arrival time after t pre is
We set T = 24 [hours] in order to consider the generated. If the time tpre is not accepted as a real
variation of online-node population during a day [5], arrival time, then the next arrival time after tpre is
[6], [15]. In terms of the PK method, the time generated. We repeat these steps throughout the
interval t[k] − t[k−1] is constant for all k’s and equal to simulation. The warm-up period is set to 500 hours
1.0 × 10−2 [hours]. We also assume that the call in order to remove the effect of the initial transient
setup processing rate of a super node is one. on simulation results. We present the mean
population size at each time in one cycle, 24 hours.
4.1 Validation of Approximation Method Figure 1 shows the mean number of ordinary
In this subsection, we validate the PK method nodes against time. In this figure, results of PK
applied in our analysis of subsection 3.3. The mean method are compared to those of simulation. We set
sojourn time of ordinary nodes, 1/μ, is 2 [hours], λ̄ = 100 and A = 50 in λ(t). This corresponds to the
and that of super nodes is 1/μ(S) = 0.02, 0.2, 2, and case where the arrival rate of ordinary nodes
20 [hours]. changes from 50 to 150 [node/hour] during a day.
In order to validate the analytical model, we In simulation results, we show the 95% confidence
conducted the simulation experiments. Throughout intervals with vertical bars.
the simulation experiments, nonstationary node- It is observed from Fig. 1 that the peak hours
arrival times are generated with the thinning computed by PK method are synchronized with the
method [10]. Let [0, t0] be a finite time interval in a phase of λ(t), while those of the simulation results
simulation execution and λmax the maximum of λ(t) get behind λ(t). Note that in PK method, the system
for 0 ≤ t ≤ t 0 . In our case, λ max = λ̄ + A. The is assumed to be stationary in piece-wise time
sequence of the arrival times is generated as intervals. Note also that the time interval t[k] − t[k−1]
follows: First, we generate “preliminary” arrival is set to 1.0 × 10−2 [hours]. This small time interval
time tpre according to the stationary Poisson process results in synchronization of λ(t) and the mean

4
Ubiquitous Computing and Communication Journal
number of user nodes. When 1/μ(S) = 0.2, 2, and 20 4000
Mean Sojourn Time of Super Nodes: 20 hours
[hours], the maximum (minimum) number of user 3500
Mean Sojourn Time of Super Nodes: 2 hours
Mean Sojourn Time of Super Nodes: 0.2 hours
nodes in PK method is almost the same as that in Mean Sojourn Time of Super Nodes: 0.02 hours
simulation. When 1/μ(S) = 0.02 [hours], however,

Mean Number of User Nodes


3000

we observe a great difference between them. Note 2500

that in PK method, the mean number of user nodes 2000


for [t[k−1], t[k]) is computed independently of the
other time intervals. If α is large in comparison with 1500

the size of node population in the system, the 1000

nonstationarity of the model greatly affects the 500


results.
Figure 2 shows the mean number of user nodes 0
0 4 8 12 16 20 24
against time for λ̄ = 1000 and A = 500. This Time (hours)
corresponds to the case where the arrival rate of
ordinary nodes changes from 500 to 1500 Figure 3: Number of user nodes vs. time. (α = 100,
[node/hour] during a day. In Fig. 2, we observe the 1/μ = 2 [hours])
same tendencies as Fig. 1. A remarkable point here
40
is that the differences between PK method and Mean Sojourn Time of Super Nodes: 20 hours

simulation are small even when 1/μ(S) = 0.02 [hours].


Mean Sojourn Time of Super Nodes: 2 hours
35
Mean Sojourn Time of Super Nodes: 0.2 hours
This is because a small α can hardly affect the size Mean Sojourn Time of Super Nodes: 0.02 hours

Mean Number of Super Nodes


30
of node population.
From Figs. 1 and 2, we can claim that PK 25

method gives a good approximation when the 20

number of user nodes is greatly larger than α. 15

10
4.2 Impact of Super-Node Sojourn Time
In this subsection, we investigate how the 5

super-node sojourn time affects the performance 0


measures. Here, we set λ̄ = 1000 and A = 500 in 0 4 8 12 16 20 24

λ(t). This corresponds to the case where the arrival Time (hours)

rate of ordinary nodes changes from 500 to 1500 Figure 4: Number of super nodes vs. time. (α = 100,
[node/hour] during a day. The mean sojourn time of 1/μ = 2 [hours])
an ordinary node, 1/μ, is 2 [hours] and that of a
super node is 1/μ(S) = 0.02, 0.2, 2, and 20 [hours]. 100
Figure 3 shows the mean number of user nodes
against time. (Remind that user nodes consist of 98

ordinary and super nodes.) All the four curves in 96


Mean Call Setup Time

Fig. 3 are symmetric and synchronize exactly with


the sinusoidal arrival rate function λ(t). This is due 94

to the approximation by the PK method. 92


In Fig. 3, the mean number of user nodes for a
larger 1/μ(S) is greater at any time. When the sojourn 90

time of a super node is small, super nodes are likely Mean Sojourn Time of Super Nodes: 20 hours
Mean Sojourn Time of Super Nodes: 2 hours
88
to become offline. As a result, an ordinary node Mean Sojourn Time of Super Nodes: 0.2 hours

frequently changes to a super node. Once the 86


Mean Sojourn Time of Super Nodes: 0.02 hours

ordinary node becomes a super node, it is likely to 0 4 8 12 16 20 24


Time (hours)
leave the network, resulting in a small population of
user nodes. Note that the mean number of user Figure 5: Mean call setup time vs. time. (α = 100,
nodes for 1/μ(S) = 20 hours is almost the same as 1/μ = 2 [hours], νS = 1)
that for 1/μ(S) = 2 hours, and that the number of user
nodes significantly decreases when 1/μ(S) is 0.02
hours. This implies that user-node population is Figure 5 represents the mean call setup time
significantly affected by the super-node sojourn over time. A remarkable point in Fig. 5 is that each
time when the super-node sojourn time is smaller curve fluctuates with a small period in addition to a
than that of the ordinary node. sinusoidal variation synchronized with the arrival
Figure 4 shows the variation of the mean rate λ(t). We also observe the increase in the
number of super nodes over time. Note that the amplitude of a short-period fluctuation when the
scale of the vertical axis in Fig. 4 is smaller than mean sojourn time of super nodes is short, or when
that in Fig. 3. We observe in Fig. 4 the same the arrival rate is small.
tendency as in Fig. 3.

5
Ubiquitous Computing and Communication Journal
1600 17 1000
Mean Sojourn Time of Super Nodes: 20 hours
Mean Sojourn Time of Super Nodes: 2 hours
100 Mean Sojourn Time of Super Nodes: 0.2 hours

Mean Churn Rate of Super Nodes


1500 16 Mean Sojourn Time of Super Nodes: 0.02 hours

Mean Number of Super Nodes


Mean Number of User Nodes

10
1400 15

1300 14
0.1

1200 13
0.01
User Nodes
Super Nodes
1100 12 0.001
15 15.2 15.4 15.6 15.8 16 0 4 8 12 16 20 24
Time (hours) Time (hours)

Figure 6: Number of user nodes vs. time. (α = 100, Figure 8: Mean churn rate of super nodes vs. time.
1/μ = 1/μ(S) = 2 [hours]) (α = 100, 1/μ = 2 [hours])

97 N(t)/N(S)(t) changes from 100 to 111. On the other


hand, when (N(t),N(S)(t)) moves from (10000, 100)
96.8 to (9999, 99), N(t)/N(S)(t) varies from 100 to 101.
This implies that the variation of the number of
Mean Call Setup Time

96.6 super-nodes for a small user-node population


significantly affects the call setup time.
96.4
Figure 8 represents the mean churn rate of
super nodes against time. We observe from Fig. 8
that the mean churn rate grows when the mean
96.2
number of user nodes increases. This is because
Call Setup Time
when there are many user nodes in the network, a
96
15 15.2 15.4 15.6 15.8 16
large number of super nodes exist in the network,
Time (hours)
resulting in a large churn rate. On the other hand,
when the super-node sojourn time 1/μ(S) is long, the
Figure 7: Mean call setup time vs. time. (α = 100, departure rate of super nodes is likely to be small.
1/μ = 1/μ(S) = 2 [hours], νS = 1) We observe that the mean churn rate grows
drastically with the decrease in 1/μ(S).
In order to examine the cause of this short-
period fluctuation, we closely look into the 4.3 Impact of the Maximum Number of User
variations of the mean numbers of nodes. Figure 6 Nodes Managed by a Super Node
shows the mean number of user nodes and that of In this subsection, we investigate the impact of
super nodes for time interval from 15 to 16 when the maximum number of user nodes managed by a
1/μ = 1/μ(S) = 2 [hours]. In this figure, the mean super node, α, on the performance of P2P systems.
number of super nodes is 14 at t = 15.2 and 15 at t = In our model, the value of α can be interpreted as
15.7. Note that in this time interval, the mean the maximum ratio of the number of user nodes to
number of user nodes monotonically increases from that of super nodes in P2P network. We consider
1320 to 1420. the call setup time and churn rate of super nodes.
Figure 7 shows the mean call setup time in the We calculated the mean call setup time over one
same time interval. The mean call setup time cycle, 24 hours, taking its time average. We define
increases from t = 15.2 to 15.5, and then decreases this value as the time average of the mean call setup
until t = 15.7. Note that the number of super nodes time. We also calculated the time average of the
changes from 14 to 15 around t = 15.5. These mean churn rate of super nodes with the same
observations imply that the mean call setup time is procedure. We set λ̄ = 1000 and A = 500 in λ(t).
significantly affected by the variation of super-node This corresponds to the case where the arrival rate
population. of ordinary nodes changes from 500 to 1500
When the mean sojourn time of super nodes is [node/hour] during a day. The mean sojourn time of
small, super nodes are likely to become offline, an ordinary node, 1/μ, is 2 [hours] and that of a
resulting in frequent transitions from an ordinary super node is 1/μ(S) = 0.02, 0.2, 2, and 20 [hours].
node to a super node. This makes the total number Figure 9 shows the time average of mean call
of user nodes small. Remind that the call setup time setup time against α. We observe that the call setup
defined by (8) is proportional to N(t)/N(S)(t). Now time is monotonically increasing with both α and
consider the transition of (N(t),N (S) (t)) and the 1/μ(S). The maximum number of user nodes
corresponding value of N(t)/N ( S ) (t). When managed by a super node, α, affects the call setup
(N(t),N(S)(t)) transits from (1000, 10) to (999, 9),

6
Ubiquitous Computing and Communication Journal
300 100000
Mean Sojourn Time of Super Nodes: 20 hours
Mean Sojourn Time of Super Nodes: 2 hours
250 Mean Sojourn Time of Super Nodes: 0.2 hours
Mean Sojourn Time of Super Nodes: 0.02 hours 10000
Mean Call Setup Time

Mean Call Setup Time


200
1000

150

100
100 P2P
Client-Server, Processing Capacity of Server 1x
10 Client-Server, Processing Capacity of Server 10x
50 Client-Server, Processing Capacity of Server 50x
Client-Server, Processing Capacity of Server 100x
Client-Server, Processing Capacity of Server 150x
0 1
0 50 100 150 200 250 300 0 4 8 12 16 20 24
Maxumum Number of User Nodes Managed by a Super Node Time (hours)

Figure 9: Time average of mean call setup time vs. Figure 11: Mean call setup time vs. time. (α = 100,
maximum number of user nodes managed by a 1/μ = 1/μ(S) = 2 [hours])
super node. (1/μ = 2 [hours], νS = 1)
4.4 Comparison of P2P and Client-Server
100 Models
In this subsection, we compare the performance
of the P2P-based user management system with that
Mean Churn Rate of Super Nodes

10

of a client-server architecture where online nodes


1 are managed by a centralized server. Suppose that
arrivals of client nodes to the network follow a
0.1
nonstationary Poisson process with rate λ(t) (t ≥ 0).
The sojourn times of online nodes are i.i.d.
according to an exponential distribution with mean
0.01 Mean Sojourn Time of Super Nodes: 20 hours
Mean Sojourn Time of Super Nodes: 2 hours 1/μ. In this case, the probability distribution of the
Mean Sojourn Time of Super Nodes: 0.2 hours number of online nodes satisfies (2) and (3) with
0.001 Mean Sojourn Time of Super Nodes: 0.02 hours
10 100
μ(S) = μ. Note that the P2P-based system can be
Maxumum Number of User Nodes Managed by a Super Node compared with the client-server-based one only
when μ(S) = μ.
Figure 10: Time average of mean churn rate of
Let νCS denote the call setup processing rate of
super nodes. (1/μ = 2 [hours])
the centralized server. Similarly to (8), we assume
that the call setup time TCS(t) is given by
time because when α is large, the number of user
nodes that a super node have to manage increases,
TCS(t) = N(t) νCS.
resulting in a large call setup time. The mean
sojourn time of super nodes, 1/μ(S), affects the call
setup time because when 1/μ(S) is small, the number For t ∈ (t[k−1], t[k]] (k = 1, 2, . . .), the mean call setup
of user nodes in the network decreases, as we time E[TCS(t)] is obtained by
observed in Fig. 3. Note that the call setup time is
significantly affected by the number of user nodes, 1 ∞
1 ∞
E[TCS (t )] = ∑ n × π n (t ) ≈ ∑n×π
[k ]
as we discussed in Fig. 7. This implies that the n .
νCS n =1 νCS n =1
super-node sojourn time affects the call setup time.
Figure 10 represents the time average of mean
We set λ̄ = 5000 and A = 2500 in λ(t). This
churn rate of super nodes against α in log-log plot.
corresponds to the case where the arrival rate of
We observe that for any 1/μ(S), the mean churn rate
ordinary nodes changes from 2500 to 7500
of super nodes decreases monotonically with the
[node/hour]. We also set 1/μ = 1/μ(S) = 2 [hours].
increase in α. This is simply because the number of
Figure 6 shows the mean call setup times for
super nodes decreases with the increase in α. We
P2P-based and client-server-based systems. In this
also observe that when 1/μ(S) is small, the churn rate
figure, νS = 1, and νCS is set to 1, 10, 50, 100, and
is significantly low in comparison with the effect of
150. In Figure 6, the mean call setup time for the
the decrement in α. This implies that a user node
P2P-based user-management system remains
with a large sojourn time should be promoted to a
constant against time, while that for the client-
super node in order to keep the churn rate of super
server-based system varies over time and is
nodes small.
synchronized with the arrival rate function λ(t), as
expected. It is also observed that the mean call
setup time for the client-server-based system
decreases with the increase in νCS.

7
Ubiquitous Computing and Communication Journal
Note that the centralized server with νCS = 150 6 REFERENCES
guarantees a call setup time smaller than or equal to
that for the P2P-based management system. In other [1] K. A. Alnowibet and H. Perros: The
words, when we use a client-server oriented Nonstationary Loss Queue: a Survey, in
network and attempt to achieve the same Modelling of Computer Systems and Networks,
performance as a P2P-based user management J. Barria, Ed., Imperial College Press (2005).
system, we require a centralized server whose [2] S. A. Baset, H. G. Schulzrinne, An Analysis of
processing rate of a call setup is 150 times faster the Skype Peer-to-Peer Internet Telephony
than that of one super node for the P2P-based Protocol, in Proc. IEEE INFOCOM 2006,
system. pp. 1-11 (2006).
From the above discussion, we can state that [3] K. Chen, C. Huang, P. Huang, and C. Lei:
P2P networks are better in signaling of real-time Quantifying Skype User Satisfaction, in Proc.
communication services in two points: we can ACM SIGCOMM 2006, pp. 399-410 (2006).
utilize nodes with a low capacity as super nodes, [4] D. P. Gaver, P. A. Jacobs, and G. Latouche:
and the QoS for the P2P-based user management Finite Birth-and-Death Models in Randomly
system is more stable than client-server oriented Changing Environments, Advances in Applied
networks. Probability, Vol. 16, pp. 715-731 (1984).
[5] L. Green, P. Kolesar, and A. Svoronos: Some
5 CONCLUSIONS Effects of Nonstationarity on Multiserver
Markovian Queueing Systems, Operations
In this paper, we have considered the Research, Vol. 39, No. 3, pp. 502-511 (1991).
performance of the decentralized management [6] S. Guha, N. Daswani, and R. Jain: An
system of user information. In our model, we have Experimental Study of the Skype Peer-to-Peer
assumed that new nodes join according to a VoIP System, in Proc. IPTPS 2006 (2006).
nonstationary Poisson process, analyzing the [7] L. Jun, Z. Shunyi, Z. Zailong, and L. Sidong:
online-node process associated with the number of Analyzing and Optimizing Skype Peer-to-Peer
super nodes. The PK method has been applied to System, in Proc. WiCOM 2007, pp. 2837-2840
the numerical calculation of performance measures. (2007).
From numerical examples, it was confirmed [8] G. Lisha and L. Junzhou: Performance
that the P2P-based user management system Analysis of a P2P-based VoIP Software, in
provides better performance than the client-server- Proc. AICT/ICIW 2006, p. 11 (2006).
based one. However, the mean call setup time of [9] E. K. Lua, J. Crowcroft, M. Pias, R. Sharma,
the P2P-based system is significantly affected by and S. Lim: A Survey and Comparison of Peer-
the mean sojourn time of super nodes. If an end- to-Peer Overlay Network Schemes, IEEE
user dislikes being a super node, he/she will Communications Surveys & Tutorials, Vol. 7,
terminate Skype application upon being a super No. 2, pp. 72-93 (2005).
node. In this case, the mean sojourn time of super [10] P. A. W. Lewis and G. S. Shedler, Simulation
nodes is likely to be short. This suggests that some of Nonhomogeneous Poisson Processes by
incentive mechanism is required in order to Thinning, Naval Research Logistics Quarterly,
guarantee stable QoS over the network. Vol. 26, No. 3, pp. 403-413 (1979).
In our analytical model, only the effect of call [11] W. A. Massey and W. Whitt: Stationary-
setup processing was taken into consideration. This Process Approximation for the Nonstationary
is because Skype is not an open protocol, and the Erlang Loss Model, Operations Research,
management mechanism for lookup is unknown. Vol. 44, No. 6, pp. 976-983 (1996).
Therefore, we considered a simple case in which [12] H. Schulzrinne and J. Rosenberg: The Session
the call setup time is inversely proportional to the Initiation Protocol: Internet-Centric Signaling,
ratio of the number of super nodes to that of user IEEE Communications Magazine, Vol. 38,
nodes. In general, however, the lookup processing No. 10, pp. 134-141 (2000).
time increases when the number of super nodes [13] H. Xie and Y. R. Yang: A Measurement-Based
grows. For future work, an analytical model taking Study of the Skype Peer-to-Peer VoIP
into account both call setup and lookup procedures Performance, in Proc. IPTPS 2007 (2007).
needs to be developed. [14] Skype, http://www.skype.com/ .
[15] Skype Statistics, http://share.skype.com/
stats_rss.xml.

8
Ubiquitous Computing and Communication Journal

You might also like