Second International Conference on Computer and Network Technology

A Framework for Performance Analysis Of

Client/Server Based SOA and P2P SOA
Piyali De, P. Chodhury, S. Choudhury
National Institute of Technology, Durgapur
M.G. Avenue, Durgapur, West Bengal, India
[email protected]
[email protected]

Abstract— An analytical framework is developed to derive the indicator in a SOA implementation. However, in [8] no
response time and service availability of client/server based effort was made to provide the necessary analytical
SOA and P2P based SOA. The impact on the response time framework so that the claims can be analytically verified.
and service availability for varying load conditions and Service availability, though being a very crucial performance
connectivity for both client/server and P2P SOA metric, was also not analyzed in any of the works mentioned
implementation is studied. It is found that the response time above.
reduces significantly in P2P based SOA depending on network In our previous work [9] a peer to peer based architecture
connectivity and service replication factors. Service availability for service oriented computing was presented, where it was
is also found to be much higher in case of P2P based SOA and
shown through emulation studies that significant
the gain are enhanced for higher values of load. In our analysis
the impacts of search latency and bandwidth penalty resulting
performance improvement can be achieved by distributing
from UDDI distribution are not yet considered. the UDDI across the peers. But the results were unable to
provide enough insight for designing efficient distributed
Keywords-SOA; Service Replication; P2P; response time; algorithms for UDDI distribution and service replication.
service availability. Moreover, effect of service replication (possibly the most
important feature of P2P system that accounts for
performance gain), was not discussed in that paper. For
I. INTRODUCTION
design and implementation of P2P-SOA, we need to design
Service oriented computing [1] is increasingly being efficient algorithms for UDDI distribution, service
adopted in enterprise computing environment for its replication, service migration and service composition.
simplicity, reusability and flexibility in new service We found it difficult to design such algorithms without
composition. Recently its popularity is extending well having a general analytical framework for estimating the
beyond the enterprise environment and global service grids crucial performance metrics. In this paper we present a
are also being designed based on Service Oriented systematic analytical framework based on which the service
Architecture (SOA) [2]. However, the intrinsic dependence response time and service availability of a SOA
on client/server based design, the technology is not well implementation can be analyzed. We consider a general
suited [3] in pervasive [4] environment. This is mainly network with average connectivity k and then we apply
because of frequent service unavailability resulting from stochastic models to such network topology to derive the
poor channel conditions, power saving strategies and user response time and availability.
mobility. The paper is organized as follows. In section II, we give a
In this paper, we develop an analytical framework for brief outline of the analytical frameworks along with the
computing the performance of client/server based SOA (C/S- assumptions we maid. We present the details of our
SOA) and peer-to-peer based SOA (P2P-SOA). Based on the derivations of response time in C/S-SOA and P2P-SOA in
framework, developed here, we compare the performance of section III. Method for evaluation of the service availability
the aforesaid architectures in terms of service response time is presented in section IV. In section V, we evaluate the
and service availability. It has been demonstrated through mathematical expression for number of nodes that are j
our analysis that the service response-time and service
availability improve significantly in P2P-SOA. away from a reference node and maximum number of hops
P2P based architecture for service oriented computing in the network. We compare the performances of C/S-SOA
was presented in [5], [6] and an agent based P2P SOA was and P2P-SOA under identical inputs and network
proposed in [7]. But none of the papers provide any configuration in the section VI. We draw our conclusions in
analytical framework for performance evaluation and hence section VII with a brief outline of the future scopes.
quantitative estimation of performance gain is not available
as far as our knowledge goes. It was shown in [8] that
response time is possibly the most effective performance

978-0-7695-4042-9/10 $26.00 © 2010 IEEE 79

DOI 10.1109/ICCNT.2010.60

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on August 17,2010 at 09:13:46 UTC from IEEE Xplore. Restrictions apply.
 II. OUT LINE OF THE ANALYTICAL FRAMEWORK time the request spends in the server including processing
We consider a wireless network with average time,
connectivity k as our reference network. By average
connectivity, in this context, we mean that on an average a 1 1
TSERVER = + . (3)
node is connected with k number of neighbors. We also μ μ −λ
assume that the network uses a symmetric MAC layer such
as 802.11. It may be noted that the logical and physical In the case of C/S-SOA the delay in the server is exactly
neighborhood becomes equivalent in wireless environment. same as in equation (3). But for P2P-SOA, with replication
factor R , the average request arrival for a particular server
We assume that there are N number of nodes in the
will be modified to (λ R ) . Hence delay in this case is
network out of which N s number of nodes act as servers that
provides unique services. We assume that servers can
consume services as well i.e. they can act both as client and 1 1
TSERVER = + . (4)
server. Let each client generates 'r ' number of requests per μ λ
μ−
unit time for each of the server. Then total number of R
requests generated by a client is given by

t = Nsr (1) B. Evaluation of TT − NETWORK

Requests from clients propagate through many hops. In
Now, it should be noted that for P2P-SOA, the network each hop it suffers a random amount of queuing delay and
consists of N s number of unique services but these services fixed amount of propagation delay. The fixed propagation
are provided by many replicated peers. Let, in the P2P delay is same for both client/server and P2P traffic and hence
scenario, the replication factor (the number of servers that will not be considered in further discussions. If ‘ h ’ is the
replicate a single service) be R . Then total number of servers average number of hops in one way path from client to
in the network becomes N s ∗ R . But the total request server, then
generated by a single client still remains the same as client-
server based system i.e. N s r . TT − NETWORK = 2 ∗ h ∗ (TPROP + TNETWORK −QUEUE ) . (5)

III. DERIVATION OF THE SERVICE RESPONSE TIME 1) Evaluation of Queuing Delay

Response time is the elapsed time or delay between We consider a large network so that the number of corner
sending request and receiving the response by the sender of nodes is significantly less than number of internal nodes so
the request i.e. service-client. Note that the response time we that we can assume uniform traffic in all network links. To
consider here is only the initial response to the service and find the queuing delay we have to find incident traffic load in
not service completion time. This response time can again be each link λl . We assume uniform load in each link
expressed as sum two components
1) Delay in the server: Delay for queuing of the service
requests that are waiting in the service queue at the
server.
2) Transmission time in the network: The round trip time
in the communication route between the client and the
server.
Figure 1. Example of a Reference Network
TRESPONSE = TSERVER + TT − NETWORK . (2) Without loss of generality, we consider X as our
reference node in figure 1. Traffic may come to node X
In the subsequent subsections, we evaluate mathematical along any of the k (4 in the figure 1) incoming links i.e.
expressions for TSERVER and TT −NETWORK . through AX, BX, CX and DX. We also assume that a client
generates request with uniform probabilities for all services.
A. Evaluation of TSERVER Let λX denotes incoming traffic into node X. If
Let the request arrival is Poisson’s process with average
i
λ X denote the amount of traffic contributed by the ith node
arrival rate λ and the server serves at a rate μ request per to λX . Then we can write,
second. Let us assume that the service time also follows
Poisson’s or deterministic distribution. Then the amount of N
λX = ∑ i λX . (6)
i =1

80

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on August 17,2010 at 09:13:46 UTC from IEEE Xplore. Restrictions apply.
 Again from observation it can be ascertained that if a replication factor R , the whole network will be divided into
1 R number of small clusters as the servers are uniformly
node i is j hops away from the reference node X then j th distributed in the network. So in this case if the whole
k
fraction of total traffic generated by i will be passing network contains N number of nodes then each small
through node X. Thus N
cluster will contain on an average number of nodes.
R
⎛ 1 ⎞ Under this situation, the parameter max hop and h will be
i
λX = N s r ∗ ⎜ ⎟.
j
⎝k ⎠ reduced resulting in reduction of λl .
Let us denote the modified max hop , h and λl for P2P-
Now, if N j denotes the number of nodes that are j hops SOA as, max hopP , hP and λlP respectively. So, for P2P-
away from X, then SOA we obtain
max hop
⎛ 1 ⎞
λX = ∑ N r ∗ N ⎜⎝ k
s j j
⎟.
⎠
(7) ⎛ 1
TT − NETWORK = 2 ∗ hP ∗ ⎜⎜
⎞
⎟⎟ . (13)
⎝ C − λlP
j =1
⎠
Where max hop denotes maximum number of hops Finally, we list out the derived response time for C/S-
between any two nodes in the network. SOA and P2P-SOA respectively as
Hence traffic through any link directly connected to X,
1 1 ⎛ 1 ⎞
λX C / S − SOA
T RESPONSE = + + 2 ∗ h ∗ ⎜⎜ ⎟.
λl = . (8) μ μ −λ ⎟
⎝ C − λl ⎠
k

Putting the value of λX from “(7)” in the above equation

we get 1 1 ⎛ 1 ⎞
P 2 P − SOA
TRESPONSE = + + 2 ∗ hP * ⎜⎜ ⎟
⎟
μ μ −λ R ⎝ C − λlP ⎠
max hop
Nj
λl = N S ∗ r ∑
j =1 k j +1
. (9)
IV. EVALUATION OF THE SERVICE AVAILABILITY
If the link capacity is constant C for every link then the The service availability in SOA depends on two factors –
queuing delay in a link can be obtained from Poisson’s link failure probability and the number of links the request
formula as and response should travel. A service may also be
unavailable to the client when the response time exceeds
1 client’s time-out setting. This issue has already been
TNETWORK −QUEUE = . (10) addressed in the previous section and we have observed that
C − λl response time gets significantly reduced in a P2P-SOA. Here
consider the effect of link failure probability in a
To determine average queuing delay in the network, we client/server and P2P environment.
have to evaluate average number of hops between any client Let Pf be the probability that a link fails, then probability
and server. It can be determined as follows
that the service is available, Pa for C/S-SOA will be given
Pa = (1 − Pf ) . And for P2P-SOA the expression
C / S − SOA h
⎛ max hop ⎞ by
⎜⎜ ∑ j ∗N j ⎟⎟
remains the same but h will be replaced with h p .
h=⎝ ⎠.
j =1
(11)
N
V. EVALUATION OF N j AND max hop
Now, putting the values of h, , TNETWORK −QUEUE in “(5)” from
Referring to the figure 1 we observed that if nodes are
the “(11)” and “(10)” respectively, we get uniformly distributed on a plain within square area and if
antennas are all having circular radiation pattern, then
⎛ 1 ⎞ connectivity of the network will be definitely be a multiple
TT − NETWORK = 2 ∗ h ∗ ⎜⎜ ⎟.
⎟ (12) of 4. This happens because of two dimensional symmetry of
⎝ C − λl ⎠ the structure. However all multiple of 4 are not valid. Now
we consider two special cases referring to the central node X.
For C/S-SOA total transmission time in the network (TT- a) First, we consider the cases where all neighbors of a
NETWORK) is same as “(12)”. But for P2P based SOA with reference node lie on or within a circle centered at the

81

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on August 17,2010 at 09:13:46 UTC from IEEE Xplore. Restrictions apply.
 reference node [e.g. k=4, 12]. In such cases the
maximum number of hops in the network can be
obtained by finding minimum value of M such that

⎛ M
⎞
⎜⎜1 + 2∑ N ∗ j ⎟⎟ ≥ N . (14)
⎝ j =1 ⎠

*
Values of N ∗ j to be used in this case N J = k ∗ j (15)
After finding the M min from “(14)” and “(15)” Figure 2. Network Delay for k=4
the max hop will be given by max hop = 2 ∗ M min
In figure 4 and figure 5, we plot the over all response
Using the value max hop we can find the values of N j as time under replication factor 4 for k=4 and k=8 respectively.
We significantly observe the response time reduces
⎧ max hop significantly in both cases for P2P SOA. As expected, here
⎪⎪k ∗ j , for j ≤ 2
also the gain in response time is enhanced with increase in
Nj = ⎨ . load.
⎪k ∗ (max hop − j + 1), for j > max hop
⎪⎩ 2

b) When all the neighboring nodes of a reference node lie

within a square centered at the reference node[e.g.
⎢ N ⎥
k=8,24], then max hop = ⎢ ⎥
⎣⎢ k + 1 − 1 ⎦⎥
Let S be an infinite sequence such that Si = 2 ∗ i ∗ 4 .
Then if the horizontal/vertical diameter of the sequence Figure 3. Network Delay for k=8
extends d nodes (not hops) excluding the reference node,
then N j will be given by Moreover, the gain is emphasized considerably for
network with higher connectivity. Here the gain in network
with k=8 is 3 times higher than that for k=4.
jd

Nj = ∑S .
( )
i = j −1 d +1
i

For other values of k also, we observed some interesting

patterns and the solution are mainly in the domain of
computational geometry. Here our objective is to analyze
some sample structures to understand the performance of
P2P system. Hence details regarding other structures are out
of scope.
VI. RESULTS AND DICURSSIONS Figure 4. Response Time for k=4
We consider a network with 256 nodes placed uniform over This emphasized behavior is solely contributed by reduction
a plain. All the nodes are having radiation pattern as in network delay as the replication factor in both the cases is
discussed above. In figure 2 and figure 3 we show the same.
network delay component of the response time against We plot service availability values against link failure
varying load for connectivity 4 and connectivity 8 probability in figure 6 and figure 7 for k=4 and k=8
respectively For P2P -SOA, we consider a replication factor respectively. We observe that while the value of service
of 4. It may be noted from the graphs that the network delay availability reduces exponentially with link failure
reduces considerably in P2P-SOA for k=4 and k=8. The probability in case of C/S-SOA, the reduction in availability
reduction in delay is emphasized for higher values of for P2P- SOA is much more slower and for k=8, this change
replication factor. The results obtained here reveal that a becomes almost linear for P2P-SOA. This shows that for
P2P-SOA gives reduced network delay not only because of higher connectivity the impact of link failure probability is
reduction of effective number of hops due to service significantly lower for P2P-SOA.
replication but also due to non-linearity in link buffer
occupancy.

82

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on August 17,2010 at 09:13:46 UTC from IEEE Xplore. Restrictions apply.
 of SOA implementation is studied. From the above studies
we can conclude that-
1) The response time reduces significantly for P2P-SOA.
This reduction is emphasized for higher network traffic
load.
2) The gain in response time in P2P-SOA is not only
because of sharing of server load but also because of
effective reduction in per link network traffic.
3) Service availability is much higher in P2P-SOA and it
Figure 5. Response Time for k=8 reduces more slowly with increase in link failure
probability in case of P2P-SOA.
It may be noted that gain in performance is achieved at
the cost of additional burden on processing, memory and
messaging overhead. This is not quantified in our analysis. It
may also be mentioned that a context aware system will
reduce the additional burden but quantitative modeling of
this is yet to be done. These two aspects should be studied in
future for designing a commercial viable system.
REFERENCES
[1] M. P. Papazoglou, P. Traverso, S. Dustdar and F. Leymann,
Figure 6. Service Availability for k=4 “Service-Oriented Computing state of the art and research challenge,”
IEEE Computer, vol. 40 , pp 38-45, November 2007.
From these result we can conclude that a significant gain [2] M. P. Papazoglou, “Service-Oriented Computing: concept,
in service response time and service availability is achieved characteristics and direction,” In Proc.,WISE 2003, 10-20 december,
when we use P2P-SOA against C/S-SOA. However this pp 3-12 .
improvement in performance is derived at the cost of [3] Geert, Vanderhulst, Kris, Luyten, Karin and Coninx, “Middleware for
additional processing and memory burden on each of the Ubiquitous service-oriented spaces on the web,” AINAW 2007, vol.
2, 21-23 May 2007, pp 1001-1006.
[4] U. Hansmann, L. Mark, M. S. Nicklous, Th. Stober, “Pervasive
Computing-The Mobile World-2nd ed.,” springer professional
series,2003.
[5] F .B. Goncalves, C. E. T. Oliveira, I. Silva, L. G. L. Moura, M. G.
Felipe, “A Software architecture for the provisioning of Mobile
service in peer-to-peer environments,” In Proc.ICIW 2007, 13-19
May 2007
[6] R. J. Bayardo, A. Crainiceanu, R. Agawal, “Peer-to-Peer sharing of
web applications,” In Proc., WWW 2003, Budapest, Hungary, May
2003
[7] Chao-Lin Wu, Chun-Feng Liao, Li-Chen Fu, ”Service-Oriented
Figure 7. Service Availability for k=8 Smart-Home architecture based on OSGI and Mobile-
Agenttechnology,” IEEE Transac. On System, Man and Cybernatics,
participating peer. For a pervasive environment dominated part C:application and review, vol.37, Issue 2, March 2007, pp-193-
by devices with little processing power and memory, simply 205.
service replication may not be a feasible solution. [8] J. S. Her, S. W. Choi, S. H. Oh, and S. D. Kim, “A framework for
measuring performance in Service-Oriented Achitecture,” In Proc.
VII. CONCLUSION AND FUTURE SCOPE Third International Conference on Next Generation Web Service
Practice, wishington, DC, USA, pp. 55-60, October 2009.
In this paper, an analytical framework is developed to [9] S. Choudury, D. Mukharjee, P. Chakraborty, N. C. Debnath and H. K.
derive the response time and service availability of C/S-SOA Lee, “An Agent based framework for peer to peer Distributed
and P2P-SOA. The impact on the response time and service Computing,” In Proc. ISCA-CATA 2008, Cancun,Mexico, pp 252-
availability for varying load and connectivity for both types 257, April 2008.

83

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY DURGAPUR. Downloaded on August 17,2010 at 09:13:46 UTC from IEEE Xplore. Restrictions apply.