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 differencedifferential 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.

INTRODUCTION

Currently existing Internet telephony service is


based on a client-server architecture with the
Session Initiation Protocol (SIP) [12]. In order to
establish a voice connection between a pair of
nodes, SIP employs three types of servers: registrar,
proxy and redirect servers. A main role of those
servers is the IP address resolution for a call
establishment. After establishing the voice
connection, voice data is exchanged directly
between the caller and callees end hosts. Although
SIP is known as a lightweight protocol, it is less
scalable and less reliable with the increase in the
number of end users due to the client-server
architecture.
Recently, Skype has been receiving
considerable attention as a peer-to-peer (P2P) VoIP
application. There are more than 370 million users
and there are at most approximately 14 million
online users [14], [15]. The services offered by
Skype are VoIP, instant messaging (IM), and filetransfer [6]. In this paper, we focus on VoIP service
of Skype.
The Skype network consists of three
components: Skype login server, ordinary node
(Skype client), and super node. The Skype login
server is a centralized server which is accessed by
1
Ubiquitous Computing and Communication Journal

ordinary nodes when joining the network. Except


the login process, Skype services are provided over
a hierarchical P2P network consisting of super
nodes and ordinary ones. Super nodes are chosen
from among ordinary nodes to manage online-node
information such as IP address and port number in a
decentralized manner. When an ordinary node
issues a voice connection request, its lookup
process is handled by the associated super nodes.
The number of super nodes changes dynamically
according to the number of online Skype nodes, and
this makes the Skype network more scalable and
more reliable than SIP-based VoIP networks.
In this paper, we model and study the
performance of this decentralized management
system of user information. Skype statistics [15]
reported that as of January, 2008, the number of
online nodes ranges from 6 million to 14 million for
a day. In order to take into account this online-node
dynamics, we assume that new nodes join the
system according to a nonstationary Poisson
process. Then we analyze the online-node process
associated with the number of super nodes with a
nonstationary Markov chain. We derive the system
of difference-differential equations for the timedependent probability distribution of the number of
online nodes to compute performance measures
using the stationary peakedness approximation

method [11]. We consider the call setup time and


churn rate of super nodes as a measure of quality of
service (QoS), and evaluate the effect of the timedependent node-join process on the call setup time.
We also compare the performance of the
decentralized management system with that of a
client-server-based system.
This paper is organized as follows. Section 2
summarizes the related work for Skype
performance and nonstationary queueing analysis.
In section 3, we describe the analytical model of
user information management in Skype, analyzing a
nonstationary Markov chain. We show some
numerical examples in section 4. Finally, section 5
concludes the paper.
2

RELATED WORK

Recently, Skype has attracted much interest not


only in industries but also in research communities.
Because Skype is not an open protocol, recent
research focuses on revealing how Skype services
are provided to millions of Skype clients. Baset and
Schulzrinne [2] reported how Skype works in the
Skype network consisting of the login server,
ordinary hosts and super nodes. From the analysis
of Skype network traffic, they focused on the
following key functions: login process, user search,
call establishment, media transfer, codecs, and
video conference.
Guha et al. [6] investigated the characteristics
of VoIP traffic in Skype by a measurement-based
analysis, comparing the traffic characteristics in
P2P file-sharing networks with that in traditional
voice-communication networks. It was reported that
Skype users are likely to run the Skype client
during office hours, causing the traffic dynamics
significantly different from that observed in P2P
file-sharing service networks. However, regardless
of these studies, network operation of each super
node has not been sufficiently revealed.
The QoS of Skype was quantitatively analyzed
by Chen et al. [3]. They proposed the User
Satisfaction Index (USI), a perceptual index for
quantifying Skype user satisfaction. USI is based on
the Cox proportional hazard model, defined with
the bitrate, jitter and Round Trip Times (RTTs)
which are obtained from a two-level sampling
approach. It was reported that user satisfaction is
significantly affected by sender transmission speed
and the superposed effect of the delay, jitter, and
packet loss.
Lisha and Junzhou [8] compared QoS of Skype
with that of MSN Messenger through a
measurement-based analysis with respect to the
refresh rate of software, connection setup delay,
one-way conversation delay and Mean Opinion
Score (MOS). They claimed that QoS of Skype is
well provided but that it is almost the same as QoS
of MSN Messenger.
2
Ubiquitous Computing and Communication Journal

Xie and Yang [13] investigated the impacts of


the access capacity and routing policy of
autonomous systems (ASes) on the QoS of Skype.
They collected the statistics of Skype super nodes
and then measured the latency and loss between
ordinary nodes and super nodes, quantifying the
Skype QoS by MOS.
Jun et al. [7] proposed two algorithms to
improve Skype QoS with analytical approach. They
focused on the user-search delay influenced with
the number of hops in a super-node overlay
network. They leveraged graph-theoretical results
for the geographical distance and relationship
proximity of small-world networks. However, there
has been little work on the fundamental analysis of
VoIP QoS over P2P-based networks.
In
general,
queueing
systems
with
nonstationary arrival processes are difficult to
analyze. Therefore, a number of approximation
methods have been proposed and studied. The
review on approximation methods for nonstationary
loss queues was provided in [1]. The effect of
nonstationary arrivals on the performance of
multiserver queueing systems was numerically
studied in [5].
Among these approximation methods is the
stationary peakedness (PK) approximation [11],
where the time axis is divided into subintervals and
the arrival process is assumed to be stationary for
each subinterval. In other words, the arrival rate at
each subinterval is regarded as a constant equal to
the time average of the arrival rate during the
subinterval. With this value, performance measures
can be calculated by analyzing a queueing system
with a stationary arrival process with the same
arrival rate.
3

ANALYSIS
OF
P2P-BASED
MANAGEMENT SYSTEM

USER

In this section, ordinary and super nodes


operations in Skype are briefly summarized, and
then the online-node process is analyzed with a
nonstationary Markov chain.
3.1

Node Operation in Skype


Recent experimental studies [2], [6] reported
how Skype application works through ordinary and
super nodes. In the following, we briefly
summarize Skypes operation for call setup and
super node promotion. The readers are referred to
[2], [6] for details.
Consider a VoIP connection establishment
between a caller and a callee over Skype. The
ordinary node of the caller issues a VoIP
connection request to his associated super node.
Skype adopts Global Index technology to search the
IP address and port number of the callees ordinary
node. If the callees node information is retrieved
and if the VoIP request is accepted, a direct VoIP

connection is established between the two ordinary


nodes.
Skype is not an open protocol, and it is not
clear how Global Index works for super node
promotion. In [6], the authors reported that the
number of super nodes varies depending on the
number of online Skype nodes. They also inferred
from experiments that Skype super nodes seem to
be chosen from the ordinary nodes with much spare
bandwidth, which are publicly reachable.
Analytical Model
In the following, the event that an offline
ordinary node becomes online is called an ordinarynode arrival. In addition, we define the sojourn time
of an ordinary/super node as the period while the
node is online.
Noting that the number of online nodes changes
in a wide range during a day, we assume that
arrivals of ordinary nodes to Skype network follow
a nonstationary Poisson process with rate (t) (t 0).
The sojourn times of ordinary nodes are i.i.d.
according to an exponential distribution with mean
1/, while the sojourn times of super nodes are
exponentially distributed with mean 1/(S).
Because of the lack of details for super node
operation, we simply assume that each super node
manages at most ( 1) nodes, including the
super node itself. Thus the total capacity of the
network is equal to N(S)(t), where N(S)(t) denotes
the number of super nodes at time t. For any time t,
we assume that the number N(S)(t) of super nodes is
minimized to be

t[ k ]
1
( t ) dt ,
[ k 1 ] t [ k 1 ]
t t
where t[k]s are chosen in such a way that [k] > 0 for
all k = 1, 2, ....
Let 0[0] = 1 and n [0] = 0 for all n = 1, 2, .... Let
[k]
(k = 1, 2,...; n = 0, 1,...) denote an
n

[ k ] =

[k ]

approximation to n(t) for t (t[k1], t[k]], i.e.,


(0(t), 1(t), . . . , ) (0[k] , 1[k] , . . .),
t (t[k1], t[k]].

3.2

According to the PK method, we determine n[k]s


(k = 1, 2,...; n = 0, 1,...) as the solution of
0 = [ k ] 0
[k ]

[k ]

[k ]

+ 11 ,

[k ]
n 1

[k ]

0=
(
n = 1,2,,

(1)

3.3

Analysis
Note that the process {N(t); t 0} is a
nonstationary birth-and-death process. Let n(t)
(t 0; n = 0, 1,...) denote Pr[N(t) = n]. It then
follows from the above assumptions that
d
(2)
0 (t ) = (t ) 0 (t ) + 11 (t ),
dt
d
n (t ) = (t ) n 1 (t ) ( (t ) + n ) n (t ) + n +1 n +1 (t ),
dt
n = 1,2,,
(3)

where n = (n n / ) + n / ( S ) . Since N(0) =


0, 0(0) = 1 and n(0) = 0 for all n = 1, 2,....
It is a hard task to solve numerically the
differential-difference equations (2) and (3).
Therefore we adopt the PK method [11] for the
computation of n(t)s. For this purpose, we
consider a sequence of times 0 = t[0] < t[1] < t[2] < ....
We then define [k] (k = 1, 2, . . .) as

3
Ubiquitous Computing and Communication Journal

+ n +1

[k ]
n +1

(5)

which are obtained from (2) and (3) by replacing


the left hand sides with zeros, and n(t)s with
n[k]s. To solve (4) and (5), we assume that there
exists for any k = 1, 2,..., some integer n*(k) such
that for all n n*(k),
n

[k ]

= n

[k ]

n < 1.

It thus follows from (4) and (5) that for any k = 1,


2,,
n

[k ]

= 0

[k ]

[k ]
i

i =1

N ( S ) (t ) = N (t ) / .

+ n ) n

(4)
[k ]

,n = 1,2,K,

(6)

[k ]

[k ]
= 1 + i .
n =1 i =1

(7)

3.4

Performance Measures
We consider the call setup time as a QoS
measure. Let S denote the call setup processing rate
of each super node. For simplicity, we assume that
the call setup processing rate for an ordinary node
is given by N(S)(t)S/N(t) when there are N(t) user
nodes and N(S)(t) super nodes at time t. We define
the call setup time TS(t) as
TS (t ) =

N (t )
.
N (t ) S

(8)

(S )

It then follows from (1) and (8) that for t


(t[k1], t[k]] (k = 1, 2, . . .), we obtain the mean call
setup time
1 N (t )
E

S N (t ) /
1 n
1
=
n (t )
S n=1 n /
CS

E[TS (t )] =

n /
n =1

[k ]
n

450
Approximation (1/mu(S) = 20)
Approximation (1/mu(S) = 2)
Approximation (1/mu(S) = 0.2)
Approximation (1/mu(S) = 0.02)
Simulation (1/mu(S) = 20)
Simulation (1/mu(S) = 2)
Simulation (1/mu(S) = 0.2)
Simulation (1/mu(S) = 0.02)

400
350
Number of User Nodes

where n[k]s are computed by (6) and (7).


It is well known that the dynamics of peer
joining and leaving, called churn, degrades the
performance of P2P-based network systems.
Therefore, the quantitative characterization of churn
is indispensable for designing P2P-based systems.
In our network model, when a super node leaves the
P2P network, user nodes administrated by the super
node need to find an alternative super node. Note
that the number of super nodes leaving the network
per unit time is equal to the number of super nodes
arriving at the network per unit time. Focusing on
the super nodes leaving the network, we define the
churn rate of super nodes CS(t) as

300
250
200
150
100
50
0
0

16

20

24

Figure 1: The number of user nodes vs. time. ( =


100, 1/ = 2 [hours])
4500
Approximation (1/mu(S) = 20)
Approximation (1/mu(S) = 2)
Approximation (1/mu(S) = 0.2)
Approximation (1/mu(S) = 0.02)
Simulation (1/mu(S) = 20)
Simulation (1/mu(S) = 2)
Simulation (1/mu(S) = 0.2)
Simulation (1/mu(S) = 0.02)

4000

It then follows from (1) and (8) that


Number of User Nodes

3500

[k ]
n
n
= ( S ) n (t ) ( S ) n ,

n =1
n =1

12
Time (hours)

C S (t ) = ( S ) N ( S ) (t ).

N (t )
E[C S (t )] = ( S ) E

3000
2500
2000
1500
1000

[k]

where n s are computed by (6) and (7).

500
0

NUMERICAL EXAMPLES

In the following, we set = 100, that is, the


maximum number of ordinary nodes a super node
can manage is 100. We consider time-dependent
arrivals of ordinary nodes with (t) = +
Acos(2t/T ), where T is the cycle length and A(> 0)
is the amplitude of the sinusoidal arrival process [5].
We set T = 24 [hours] in order to consider the
variation of online-node population during a day [5],
[6], [15]. In terms of the PK method, the time
interval t[k] t[k1] is constant for all ks and equal to
1.0 102 [hours]. We also assume that the call
setup processing rate of a super node is one.
4.1

Validation of Approximation Method


In this subsection, we validate the PK method
applied in our analysis of subsection 3.3. The mean
sojourn time of ordinary nodes, 1/, is 2 [hours],
and that of super nodes is 1/(S) = 0.02, 0.2, 2, and
20 [hours].
In order to validate the analytical model, we
conducted the simulation experiments. Throughout
the simulation experiments, nonstationary nodearrival times are generated with the thinning
method [10]. Let [0, t0] be a finite time interval in a
simulation execution and max the maximum of (t)
for 0 t t 0 . In our case, max = + A. The
sequence of the arrival times is generated as
follows: First, we generate preliminary arrival
time tpre according to the stationary Poisson process

4
Ubiquitous Computing and Communication Journal

12

16

20

24

Time (hours)

Figure 2: The number of user nodes vs. time. ( =


100, 1/ = 2 [hours])

with rate max. Then, tpre is accepted as a real


arrival time with probability (tpre)/max. After that,
the next preliminary arrival time after t pre is
generated. If the time tpre is not accepted as a real
arrival time, then the next arrival time after tpre is
generated. We repeat these steps throughout the
simulation. The warm-up period is set to 500 hours
in order to remove the effect of the initial transient
on simulation results. We present the mean
population size at each time in one cycle, 24 hours.
Figure 1 shows the mean number of ordinary
nodes against time. In this figure, results of PK
method are compared to those of simulation. We set
= 100 and A = 50 in (t). This corresponds to the
case where the arrival rate of ordinary nodes
changes from 50 to 150 [node/hour] during a day.
In simulation results, we show the 95% confidence
intervals with vertical bars.
It is observed from Fig. 1 that the peak hours
computed by PK method are synchronized with the
phase of (t), while those of the simulation results
get behind (t). Note that in PK method, the system
is assumed to be stationary in piece-wise time
intervals. Note also that the time interval t[k] t[k1]
is set to 1.0 102 [hours]. This small time interval
results in synchronization of (t) and the mean

Impact of Super-Node Sojourn Time


In this subsection, we investigate how the
super-node sojourn time affects the performance
measures. Here, we set = 1000 and A = 500 in
(t). This corresponds to the case where the arrival
rate of ordinary nodes changes from 500 to 1500
[node/hour] during a day. The mean sojourn time of
an ordinary node, 1/, is 2 [hours] and that of a
super node is 1/(S) = 0.02, 0.2, 2, and 20 [hours].
Figure 3 shows the mean number of user nodes
against time. (Remind that user nodes consist of
ordinary and super nodes.) All the four curves in
Fig. 3 are symmetric and synchronize exactly with
the sinusoidal arrival rate function (t). This is due
to the approximation by the PK method.
In Fig. 3, the mean number of user nodes for a
larger 1/(S) is greater at any time. When the sojourn
time of a super node is small, super nodes are likely
to become offline. As a result, an ordinary node
frequently changes to a super node. Once the
ordinary node becomes a super node, it is likely to
leave the network, resulting in a small population of
user nodes. Note that the mean number of user
nodes for 1/(S) = 20 hours is almost the same as
that for 1/(S) = 2 hours, and that the number of user
nodes significantly decreases when 1/(S) is 0.02
hours. This implies that user-node population is
significantly affected by the super-node sojourn
time when the super-node sojourn time is smaller
than that of the ordinary node.
Figure 4 shows the variation of the mean
number of super nodes over time. Note that the
scale of the vertical axis in Fig. 4 is smaller than
that in Fig. 3. We observe in Fig. 4 the same
tendency as in Fig. 3.

5
Ubiquitous Computing and Communication Journal

Mean Sojourn Time of Super Nodes: 20 hours


Mean Sojourn Time of Super Nodes: 2 hours
Mean Sojourn Time of Super Nodes: 0.2 hours
Mean Sojourn Time of Super Nodes: 0.02 hours

Mean Number of User Nodes

3500
3000
2500
2000
1500
1000
500
0
0

12

16

20

24

Time (hours)

Figure 3: Number of user nodes vs. time. ( = 100,


1/ = 2 [hours])
40

Mean Sojourn Time of Super Nodes: 20 hours


Mean Sojourn Time of Super Nodes: 2 hours
Mean Sojourn Time of Super Nodes: 0.2 hours
Mean Sojourn Time of Super Nodes: 0.02 hours

35
Mean Number of Super Nodes

4.2

4000

30
25
20
15
10
5
0
0

12

16

20

24

Time (hours)

Figure 4: Number of super nodes vs. time. ( = 100,


1/ = 2 [hours])
100
98

Mean Call Setup Time

number of user nodes. When 1/(S) = 0.2, 2, and 20


[hours], the maximum (minimum) number of user
nodes in PK method is almost the same as that in
simulation. When 1/(S) = 0.02 [hours], however,
we observe a great difference between them. Note
that in PK method, the mean number of user nodes
for [t[k1], t[k]) is computed independently of the
other time intervals. If is large in comparison with
the size of node population in the system, the
nonstationarity of the model greatly affects the
results.
Figure 2 shows the mean number of user nodes
against time for = 1000 and A = 500. This
corresponds to the case where the arrival rate of
ordinary nodes changes from 500 to 1500
[node/hour] during a day. In Fig. 2, we observe the
same tendencies as Fig. 1. A remarkable point here
is that the differences between PK method and
simulation are small even when 1/(S) = 0.02 [hours].
This is because a small can hardly affect the size
of node population.
From Figs. 1 and 2, we can claim that PK
method gives a good approximation when the
number of user nodes is greatly larger than .

96
94
92
90
Mean Sojourn Time of Super Nodes: 20 hours
Mean Sojourn Time of Super Nodes: 2 hours
Mean Sojourn Time of Super Nodes: 0.2 hours
Mean Sojourn Time of Super Nodes: 0.02 hours

88
86
0

12

16

20

24

Time (hours)

Figure 5: Mean call setup time vs. time. ( = 100,


1/ = 2 [hours], S = 1)

Figure 5 represents the mean call setup time


over time. A remarkable point in Fig. 5 is that each
curve fluctuates with a small period in addition to a
sinusoidal variation synchronized with the arrival
rate (t). We also observe the increase in the
amplitude of a short-period fluctuation when the
mean sojourn time of super nodes is short, or when
the arrival rate is small.

1500

16

1400

15

1300

14

1200

13
User Nodes
Super Nodes

1100
15

15.2

15.4

15.6

1000

Mean Churn Rate of Super Nodes

17

Mean Number of Super Nodes

Mean Number of User Nodes

1600

12

15.8

96.8
Mean Call Setup Time

0.1

0.01

12

16

20

24

Time (hours)

97

96.6

96.4

96.2

Call Setup Time

96
15.4

10

Figure 6: Number of user nodes vs. time. ( = 100,


1/ = 1/(S) = 2 [hours])

15.2

100

0.001

16

Time (hours)

15

Mean Sojourn Time of Super Nodes: 20 hours


Mean Sojourn Time of Super Nodes: 2 hours
Mean Sojourn Time of Super Nodes: 0.2 hours
Mean Sojourn Time of Super Nodes: 0.02 hours

15.6

15.8

16

Time (hours)

Figure 7: Mean call setup time vs. time. ( = 100,


1/ = 1/(S) = 2 [hours], S = 1)

In order to examine the cause of this shortperiod fluctuation, we closely look into the
variations of the mean numbers of nodes. Figure 6
shows the mean number of user nodes and that of
super nodes for time interval from 15 to 16 when
1/ = 1/(S) = 2 [hours]. In this figure, the mean
number of super nodes is 14 at t = 15.2 and 15 at t =
15.7. Note that in this time interval, the mean
number of user nodes monotonically increases from
1320 to 1420.
Figure 7 shows the mean call setup time in the
same time interval. The mean call setup time
increases from t = 15.2 to 15.5, and then decreases
until t = 15.7. Note that the number of super nodes
changes from 14 to 15 around t = 15.5. These
observations imply that the mean call setup time is
significantly affected by the variation of super-node
population.
When the mean sojourn time of super nodes is
small, super nodes are likely to become offline,
resulting in frequent transitions from an ordinary
node to a super node. This makes the total number
of user nodes small. Remind that the call setup time
defined by (8) is proportional to N(t)/N(S)(t). Now
consider the transition of (N(t),N (S) (t)) and the
corresponding value of N(t)/N ( S ) (t). When
(N(t),N(S)(t)) transits from (1000, 10) to (999, 9),
6
Ubiquitous Computing and Communication Journal

Figure 8: Mean churn rate of super nodes vs. time.


( = 100, 1/ = 2 [hours])

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


hand, when (N(t),N(S)(t)) moves from (10000, 100)
to (9999, 99), N(t)/N(S)(t) varies from 100 to 101.
This implies that the variation of the number of
super-nodes for a small user-node population
significantly affects the call setup time.
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
number of user nodes increases. This is because
when there are many user nodes in the network, a
large number of super nodes exist in the network,
resulting in a large churn rate. On the other hand,
when the super-node sojourn time 1/(S) is long, the
departure rate of super nodes is likely to be small.
We observe that the mean churn rate grows
drastically with the decrease in 1/(S).
4.3

Impact of the Maximum Number of User


Nodes Managed by a Super Node
In this subsection, we investigate the impact of
the maximum number of user nodes managed by a
super node, , on the performance of P2P systems.
In our model, the value of can be interpreted as
the maximum ratio of the number of user nodes to
that of super nodes in P2P network. We consider
the call setup time and churn rate of super nodes.
We calculated the mean call setup time over one
cycle, 24 hours, taking its time average. We define
this value as the time average of the mean call setup
time. We also calculated the time average of the
mean churn rate of super nodes with the same
procedure. We set = 1000 and A = 500 in (t).
This corresponds to the case where the arrival rate
of ordinary nodes changes from 500 to 1500
[node/hour] during a day. The mean sojourn time of
an ordinary node, 1/, is 2 [hours] and that of a
super node is 1/(S) = 0.02, 0.2, 2, and 20 [hours].
Figure 9 shows the time average of mean call
setup time against . We observe that the call setup
time is monotonically increasing with both and
1/(S). The maximum number of user nodes
managed by a super node, , affects the call setup

300

100000

10000
Mean Call Setup Time

Mean Call Setup Time

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
200

150

100

1000

100
P2P
Client-Server, Processing Capacity of Server 1x
Client-Server, Processing Capacity of Server 10x
Client-Server, Processing Capacity of Server 50x
Client-Server, Processing Capacity of Server 100x
Client-Server, Processing Capacity of Server 150x

10

50

1
0

50

100

150

200

250

300

Maxumum Number of User Nodes Managed by a Super Node

Figure 9: Time average of mean call setup time vs.


maximum number of user nodes managed by a
super node. (1/ = 2 [hours], S = 1)

12

16

20

24

Time (hours)

Figure 11: Mean call setup time vs. time. ( = 100,


1/ = 1/(S) = 2 [hours])
4.4

Mean Churn Rate of Super Nodes

100

10

0.1

0.01

Mean Sojourn Time of Super Nodes: 20 hours


Mean Sojourn Time of Super Nodes: 2 hours
Mean Sojourn Time of Super Nodes: 0.2 hours
0.001 Mean Sojourn Time of Super Nodes: 0.02 hours
10
100
Maxumum Number of User Nodes Managed by a Super Node

Figure 10: Time average of mean churn rate of


super nodes. (1/ = 2 [hours])

time because when is large, the number of user


nodes that a super node have to manage increases,
resulting in a large call setup time. The mean
sojourn time of super nodes, 1/(S), affects the call
setup time because when 1/(S) is small, the number
of user nodes in the network decreases, as we
observed in Fig. 3. Note that the call setup time is
significantly affected by the number of user nodes,
as we discussed in Fig. 7. This implies that the
super-node sojourn time affects the call setup time.
Figure 10 represents the time average of mean
churn rate of super nodes against in log-log plot.
We observe that for any 1/(S), the mean churn rate
of super nodes decreases monotonically with the
increase in . This is simply because the number of
super nodes decreases with the increase in . We
also observe that when 1/(S) is small, the churn rate
is significantly low in comparison with the effect of
the decrement in . This implies that a user node
with a large sojourn time should be promoted to a
super node in order to keep the churn rate of super
nodes small.

7
Ubiquitous Computing and Communication Journal

Comparison of P2P and Client-Server


Models
In this subsection, we compare the performance
of the P2P-based user management system with that
of a client-server architecture where online nodes
are managed by a centralized server. Suppose that
arrivals of client nodes to the network follow a
nonstationary Poisson process with rate (t) (t 0).
The sojourn times of online nodes are i.i.d.
according to an exponential distribution with mean
1/. In this case, the probability distribution of the
number of online nodes satisfies (2) and (3) with
(S) = . Note that the P2P-based system can be
compared with the client-server-based one only
when (S) = .
Let CS denote the call setup processing rate of
the centralized server. Similarly to (8), we assume
that the call setup time TCS(t) is given by

TCS(t) = N(t) CS.


For t (t[k1], t[k]] (k = 1, 2, . . .), the mean call setup
time E[TCS(t)] is obtained by
E[TCS (t )] =

1
CS

n n (t )
n =1

1
CS

[k ]
n

n =1

We set = 5000 and A = 2500 in (t). This


corresponds to the case where the arrival rate of
ordinary nodes changes from 2500 to 7500
[node/hour]. We also set 1/ = 1/(S) = 2 [hours].
Figure 6 shows the mean call setup times for
P2P-based and client-server-based systems. In this
figure, S = 1, and CS is set to 1, 10, 50, 100, and
150. In Figure 6, the mean call setup time for the
P2P-based user-management system remains
constant against time, while that for the clientserver-based system varies over time and is
synchronized with the arrival rate function (t), as
expected. It is also observed that the mean call
setup time for the client-server-based system
decreases with the increase in CS.

Note that the centralized server with CS = 150


guarantees a call setup time smaller than or equal to
that for the P2P-based management system. In other
words, when we use a client-server oriented
network and attempt to achieve the same
performance as a P2P-based user management
system, we require a centralized server whose
processing rate of a call setup is 150 times faster
than that of one super node for the P2P-based
system.
From the above discussion, we can state that
P2P networks are better in signaling of real-time
communication services in two points: we can
utilize nodes with a low capacity as super nodes,
and the QoS for the P2P-based user management
system is more stable than client-server oriented
networks.
5

CONCLUSIONS

In this paper, we have considered the


performance of the decentralized management
system of user information. In our model, we have
assumed that new nodes join according to a
nonstationary Poisson process, analyzing the
online-node process associated with the number of
super nodes. The PK method has been applied to
the numerical calculation of performance measures.
From numerical examples, it was confirmed
that the P2P-based user management system
provides better performance than the client-serverbased one. However, the mean call setup time of
the P2P-based system is significantly affected by
the mean sojourn time of super nodes. If an enduser dislikes being a super node, he/she will
terminate Skype application upon being a super
node. In this case, the mean sojourn time of super
nodes is likely to be short. This suggests that some
incentive mechanism is required in order to
guarantee stable QoS over the network.
In our analytical model, only the effect of call
setup processing was taken into consideration. This
is because Skype is not an open protocol, and the
management mechanism for lookup is unknown.
Therefore, we considered a simple case in which
the call setup time is inversely proportional to the
ratio of the number of super nodes to that of user
nodes. In general, however, the lookup processing
time increases when the number of super nodes
grows. For future work, an analytical model taking
into account both call setup and lookup procedures
needs to be developed.

8
Ubiquitous Computing and Communication Journal

REFERENCES

[1] K. A. Alnowibet and H. Perros: The


Nonstationary Loss Queue: a Survey, in
Modelling of Computer Systems and Networks,
J. Barria, Ed., Imperial College Press (2005).
[2] S. A. Baset, H. G. Schulzrinne, An Analysis of
the Skype Peer-to-Peer Internet Telephony
Protocol, in Proc. IEEE INFOCOM 2006,
pp. 1-11 (2006).
[3] K. Chen, C. Huang, P. Huang, and C. Lei:
Quantifying Skype User Satisfaction, in Proc.
ACM SIGCOMM 2006, pp. 399-410 (2006).
[4] D. P. Gaver, P. A. Jacobs, and G. Latouche:
Finite Birth-and-Death Models in Randomly
Changing Environments, Advances in Applied
Probability, Vol. 16, pp. 715-731 (1984).
[5] L. Green, P. Kolesar, and A. Svoronos: Some
Effects of Nonstationarity on Multiserver
Markovian Queueing Systems, Operations
Research, Vol. 39, No. 3, pp. 502-511 (1991).
[6] S. Guha, N. Daswani, and R. Jain: An
Experimental Study of the Skype Peer-to-Peer
VoIP System, in Proc. IPTPS 2006 (2006).
[7] L. Jun, Z. Shunyi, Z. Zailong, and L. Sidong:
Analyzing and Optimizing Skype Peer-to-Peer
System, in Proc. WiCOM 2007, pp. 2837-2840
(2007).
[8] G. Lisha and L. Junzhou: Performance
Analysis of a P2P-based VoIP Software, in
Proc. AICT/ICIW 2006, p. 11 (2006).
[9] E. K. Lua, J. Crowcroft, M. Pias, R. Sharma,
and S. Lim: A Survey and Comparison of Peerto-Peer Overlay Network Schemes, IEEE
Communications Surveys & Tutorials, Vol. 7,
No. 2, pp. 72-93 (2005).
[10] P. A. W. Lewis and G. S. Shedler, Simulation
of Nonhomogeneous Poisson Processes by
Thinning, Naval Research Logistics Quarterly,
Vol. 26, No. 3, pp. 403-413 (1979).
[11] W. A. Massey and W. Whitt: StationaryProcess Approximation for the Nonstationary
Erlang Loss Model, Operations Research,
Vol. 44, No. 6, pp. 976-983 (1996).
[12] H. Schulzrinne and J. Rosenberg: The Session
Initiation Protocol: Internet-Centric Signaling,
IEEE Communications Magazine, Vol. 38,
No. 10, pp. 134-141 (2000).
[13] H. Xie and Y. R. Yang: A Measurement-Based
Study of the Skype Peer-to-Peer VoIP
Performance, in Proc. IPTPS 2007 (2007).
[14] Skype, http://www.skype.com/ .
Statistics,
http://share.skype.com/
[15] Skype
stats_rss.xml.

You might also like