International Journal of Computer Science Engineering

and Information Technology Research (IJCSEITR)

ISSN(P): 2249-6831; ISSN(E): 2249-7943
Vol. 4, Issue 6, Dec 2014, 41-50
TJPRC Pvt. Ltd.

A NOVEL ANALYSIS FOR STABILITY AND PERFORMANCE OF NETWORK

RESOURCE ALLOCATION AND NUM PROBLEM IN LARGE SCALE NETWORKS
SAI ASHWIN PARAKKAL1, PRASANNA BALASUBRAMANIAN2 & T. SARAVANAN3
1,2
3

Research Scholar, Computer Science Engineering, Bharath University, Selaiyur, Chennai, Tamil Nadu, India

HOD, Electronics and Telecommunication Engineering, Bharath University, Selaiyur, Chennai, Tamil Nadu, India

ABSTRACT
Network resource allocation controls the number of different routes, where each connection is subject to
congestion control. The fundamental problem in telecommunication network among the shared infrastructure is resource
allocation with fairness and stability. The Network Utility Maximization (NUM) problem captures various fairness notions
between end-to-end flows and takes care of congestion control. Flows or end-to-end connections in networks dynamically
share resources according to various resource allocation schemes. These flows can be identified through their classes,
which define the set of network resources they require for the transfer of the corresponding packets. An existing schemes
of allocate resources to competing flows by distributive solving a network utility maximization problem.
In this paper we studied stability of network resource allocation against large scale networks. And we propose a
decentralized admission control and primal dual congestion control based on user utilities and thus tailored to our proposed
user-centric stability. We analyze the performance of this control under a traffic model of random connection
arrival/departures through a fluid limit argument. Our result is shown to protect the network from users, imposing in
situations of overload the desired notion of stability. The important problem analyzed here is the resource allocation
provided by congestion control algorithms over a physical layer that allows multiple transmission rates, such as wired
networks. We show that the typical algorithms in use lead to important inefficiencies from the connection-level
perspective, and we propose mechanisms to overcome these inefficiencies and enhance the resource allocation provided by
such networks.

KEYWORDS: Network Utility Maximization (NUM), Communication Networks, Optimization, Stochastic Process
1. INTRODUCTION
Modern communication networks must encompass and simultaneously support multiple users, services and
applications with diverse demands and requirements that push networks performance closer to their limit. Resource
allocation for communication networks is the process of deciding how a set of network resources are used. In this
dissertation, scenarios of wireless systems contain resources in terms of time, frequency, and power. Allocating time
resources can be mainly exemplified by link scheduling. In particular, resource allocation is the process of selecting which
link(s) to be active during a specific time slot. The same time slot can be assigned to multiple links as long as they are not
generating interference to any other link. Resources in terms of frequency are limited as any communication system is
always constrained by bandwidth. Each active link necessitates a certain amount of bandwidth to be able to send data.
Frequency resources require careful allocation because two mutually interfered links cannot be active within the same
bandwidth simultaneously. Therefore, optimum resource allocation between users and/or applications is of paramount
www.tjprc.org

editor@tjprc.org

42

Sai Ashwin Parakkal, Prasanna Balasubramanian & T. Saravanan

importance in order to assure efficient utilization of the network.

The resource allocation problem is one of the numerous research areas in which Optimization Theory has found
extensive use, since it can lead to the development of distributed algorithms to assure optimal allocation of resources in a
network. An inherent characteristic of many real world networks that makes resource allocation challenging is the
decentralization of information. Information decentralization arises due to the following reasons: One, the large-scale
nature of the networks. To have a centralized control in such networks requires the communication of enormous amounts
of data to a central controller and this is practically infeasible. For example, in a sensor network with thousands of sensors,
it is difficult to keep a centralized record of the data collected at all sensors. Two, the presence of selfish/competitive
network agents who do not want to reveal their private information and thus make centralized control infeasible Kelly et al.
in their seminal paper [1], and Low et al. [2] later using a different mathematical approach, introduced the Network Utility
Maximization (NUM) framework, where the resource allocation problem is expressed as an optimization problem. Based
on different network scenarios, smart resource allocation leads to optimization of different performance objectives. In this
paper, we consider various performance objectives according to specific models. For example, throughput is the objective
that is always desirable for most kinds of large scale networks. In a network each user having number of connection
through different routers the stability and efficiency of network resource allocation between users that optimizes the
Network Utility Maximization (NUM) problem. And we propose a decentralized admission control and primal dual
congestion control based on user utilities and thus tailored to our proposed user-centric stability.
The first step in this direction came with the work of [Jacobson, 1988], in which the author proposed to add a
congestion control mechanism to the Transmission Control Protocol (TCP) which was and is used on the Internet to
transfer data files. The algorithm was based on the sliding window mechanism that was already in use for flow control, but
with the possibility to adapt the window size based on network congestion. The main assumption is that packet losses in a
given communication were caused by some resources of the network becoming congested, and thus upon detection of a
lost segment in the data transfer, the sender must restrain from sending too much data, by reducing its transmission
window. The proposed algorithm, which we will describe later, as well as its successive updates have been undoubtedly
successful in steering the traffic of the Internet through its global expansion.
The importance of the congestion control problem has attracted a large research community. The main
breakthrough was the formulation in the seminal work of [Kelly et al., 1998]. Kelly poses the congestion control problem
as a suitable convex optimization problem, where connections are represented as economic actors in a bandwidth market.
The network congestion signals are interpreted as prices, and the market equilibrium becomes the resource allocation
obtained through the decentralized mechanisms. The resulting formulation has become known in the research community
as the Network Utility Maximization (NUM) problem. Throughout this work, we shall apply the NUM framework to
analyze several important problems of network resource allocation. One of the main issues with economic models for
network congestion control is the fact that the utility a connection derives in the aforementioned bandwidth market is
associated with the protocol behavior. It is thus only a metaphorical description, not related with real user willingness for
bandwidth. In general, typical users are not aware of this, because the congestion control algorithms are deep inside the
operating systems of their hosts. However, several applications started to cheat the bandwidth allocation by opening
multiple connections.

Impact Factor (JCC): 6.8785

Index Copernicus Value (ICV): 3.0

43

A Novel Analysis for Stability and Performance of Network

Resource Allocation and Num Problem in Large Scale Networks

2. Resource Allocation and Congestion Control in Networks

We will present the main ideas behind the Network Utility Maximization (NUM) framework, which serve as a
basis for the development of the contributions of the paper. The NUM theory has evolved since the work of [Kelly et al.,
1998] and is nowadays used to model different problems in networking. Below we introduce the reader to the main results
and we establish some important definitions and notations used throughout this work. A main reference for NUM theory is
[Srikant, 2004].
2.1. Resource Allocation and Ooptimization Problem
We begin by a simple example that illustrates the relationship between resource allocation and optimization, and
provides intuition about some of the concepts involved later. Consider a set of entities, which we may call users and we
index by i. These users want to share some scarce resource. Assume that the total available quantity of the resource is
limited by some amount c > 0. If a given user i is allocated a given amount xi of the total resource, it derives a utility Ui
(xi). From now on, we make the following assumption on utility functions:
Problem 2.1
The utility function Ui () is a continuously differentiable, non decreasing, strictly concave function Ui: R+R.
The utility function expresses the benefit a user obtains from the resources. The non decreasing hypothesis states
simply that the user is happier whenever more resources are allocated to it. The concave assumption is standard, and
expresses the fact the amount of marginal utility provided by extra resources is decreasing.
A resource allocation is a vector x = (xi), with xi>0, that indicates the amount of resources given to each one of the
users. Since the total amount is limited by c, a resource allocation would be feasible if and only if:

(1)
Clearly, there are several possible resource allocations. We are interested in finding a resource allocation that
gives an optimal trade-off between all the participating entities. A reasonable choice is to maximize the social welfare,
which is given by the following convex optimization problem:

(2)
Subject to constraint:

(3)
The above problem is the optimization of a concave function, due to Assumption 2.1, subject to the linear (and
therefore convex) inequality constraint (2.1). We write the Lagrangian of the problem, which is:

(4)
www.tjprc.org

editor@tjprc.org

44

Sai Ashwin Parakkal, Prasanna Balasubramanian & T. Saravanan

Where p denotes the Lagrange multiplier associated with the constraint.

The Karush-Kuhn-Tucker optimality conditions for this problem state that any solution (x*, p*) must satisfy:

(5)
Where p*> 0, and x* satisfies the constraint (2.1).
We can discuss two cases. In the first case, optimality is achieved with strict inequality in the constraint. In this
case, the second condition above implies that p*= 0.

Figure 1: User Utility and Demand Curve

This happens when there is a point in the interior of the region defined by (2.1) where all users achieve maximal
individual satisfaction. In this case, we say that the constraint is not active and the associated Lagrange multiplier will be 0.
The simple example presented here illustrates the main concepts and intuition behind the economic interpretation of
resource allocation.
3. Congestion Control Algorithms in the Internet
The transmission of data flows over the Internet is done mainly using the Transmission Control Protocol (TCP)
[RFC 793, 1981]. This protocol provides a reliable transport over an unreliable datagram network. The protocol takes the
input flow of data from the application layer and segments it into packets, which will be delivered to the network layer
below, in charge of relaying these packets to the destination. In order to recover from possible losses in the
communication, TCP adds a sequence number to each packet and the destination sends an acknowledgment (ACK) packet
in the reverse path to tell the source the data has correctly arrived. When TCP detects a missing acknowledgment, it will
interpret it as a lost packet and retransmit the data. When this happens, the receiver must buffer the packets until it fills all
the gaps, so it can deliver an ordered stream to the application layer above at the destination.
In order to control the number of in-flight packets (i.e. packets sent but whose ACK has not yet been received),
the sender side maintains a transmission window W, which is the maximum allowed number of in-flight packets. The
simplest case is W = 1 which is the stop-and wait algorithm. In that case, the sender waits for the ACK of the previous
Impact Factor (JCC): 6.8785

Index Copernicus Value (ICV): 3.0

A Novel Analysis for Stability and Performance of Network

Resource Allocation and Num Problem in Large Scale Networks

45

packet before proceeding with the next. This choice of W is clearly inefficient when the Round Trip Time (RTT) between
source and destination is large, since it can only send one packet per RTT. A more efficient situation can be achieved by
enlarging the number of in flight packets as depicted in Figure 2. In that case, the average rate obtained by the source will
be given by:

(6)
Suppose now that the network is able to provide a maximum amount of bandwidth c (in packets per second). In
that case, an optimal choice of W will be equal to the bandwidth delay

Figure 2: Sliding Window Protocol and Rate Calculation

Product c d , with d the round-trip propagation delay of the network. Setting W greater than this value will only
increase the total RT T due to buffered packets that cannot be immediately delivered, and the average rate will still be
equal to c. Setting W lower than this value will not make use of the full capacity of the network. The problem is that, even
if the source can measure its RT T , it does not know c a priori, and most importantly, the value of this allocated capacity
may be time varying due to other connections competing for the same resources. In [Jacobson, 1988], the author described
a simple algorithm to adapt the value of W. We now describe this algorithm and show how it can be interpreted as a primal
adaptation in the language of Network Utility Maximization.
4. User-Centric Fairness
Let us assume, as before, that the network is composed by a set of resources or links of capacity cl. Routes,
indexed by r are established along these links, and we have, as before, a routing matrix Rlr which stores the incidence of
routes on links. Above this network, we have new entities which we call users, and we index them by i .Users open one or
more end to end connections through the network, over a set of routes. If a given user opens connections over route r we
write r i . Note that the same route can be in use by more than one user. In that case we duplicate the index r , indexing
routes of different users with different labels.
These connections obtain from the network an aggregate rate on route r, and we define the total rate a given user
obtains from the network through all of its routes as:

www.tjprc.org

editor@tjprc.org

46

Sai Ashwin Parakkal, Prasanna Balasubramanian & T. Saravanan

(7)
Let Ui be an increasing and concave utility function that models user preferences or desire for bandwidth,
independent of the protocols in place in the network. We propose the following notion of user-centric fairness, defined via
the following NUM problem:

(8)
Subject to:

(9)
The above Problem amounts to finding a suitable resource allocation for every route in the network so that the
total user welfare is maximized. By using different utility functions (e.g. the a-family of Definition 2.) we can model
different notions of fairness between users. Note also that the framework is very general: a user is defined by a set of routes
and a utility function. This can model for instance users downloading data from several locations, multiple parallel paths,
users uploading data to several points in the network, and of course, the single path situations discussed earlier. To see this
note that the objective function can be rewritten as:

(10)
Therefore, switching some resources from a route r i to another route r i, while maintaining the value of
does not change the value of the objective function. Thus any convex combination r + (1 - ) r produces the same total
welfare. Of course, in the single-path case where each user is confined to a single route, Problem coincide. In the following
Section we will show how to address this issue, assuming that users can directly control the rate r for each r i . Then we
will proceed to find suitable laws for nr, the number of connections in each route, that emulate the desired behavior for r.
4.1. Controlling the Number of Flows: The Single-Path Case
Based on the preceding discussions, we would like to find a suitable control law for the number of ongoing flows
in each route such that the system converges to the resource allocation defined by Problem 4.1. In this section, we will
focus on the simpler situation in which each user opens connections over only one route in the network. Therefore we can
identify a user i and its route r, and the aggregate rate obtained by the user is i=r.
We will assume also that the network operates with some suitable congestion control algorithm such that, for a
fixed number of connections n = (nr), the network seeks to optimize the TCP Congestion Control Problem. For ease of
exposition we will assume that this congestion control is a dual algorithm, but we discuss the alternatives below.
Under these assumptions, the network part of the dynamics will be given by:

Impact Factor (JCC): 6.8785

Index Copernicus Value (ICV): 3.0

A Novel Analysis for Stability and Performance of Network

Resource Allocation and Num Problem in Large Scale Networks

47

(11)
Here xr represents the individual rate of each connection on route r . Given a current route price qr, the rate is
determined by the condition:

(12)
Where UTCP,r models the TCP behavior. Therefore, fTCP,r is the demand curve for the utility function UTCPr, or U1

(qr). In the case of a-fairness, where UTCPr(xr) = wrxr-a this demand curve is given by:

(13)
and it will be a decreasing function of the price.
We propose to add, on top of this layer, a control law for n r such that in equilibrium the rate r *= nr* xr* is
determined by the user level utility function Ur instead of the network layer utility UTCPr. The main intuition is to use the
current lower layer congestion price as a feedback signal from the network. If under the current conditions the resource
allocation is such that user i with route r gets a rate r which does not fullfill the user demand, the rate must be increased.
In the light of Theorem3, we can do so by increasing nr .

Figure 3: Combined Connection Level User Level Control Loop

The above argument justifies why we can use the same route congestion price qr to control both the rate of each
connection and also the number of connections. The proposed control loop is best explained through Figure 3, where we
can see that the connection level control is an outer control loop added to the current network congestion control algorithm.

www.tjprc.org

editor@tjprc.org

48

Sai Ashwin Parakkal, Prasanna Balasubramanian & T. Saravanan

5. RESULT AND ANALYSIS

We discuss practical implementation issues and investigate the performance of the policies developed through
simulations. We do so in several scenarios. In the first one, we consider the case where users cooperate controlling their
number of active connections in a proactive way to achieve a fair resource allocation according to the User Welfare
Problem 2. Then, we move on to explore the behavior of utility-based admission control on an overloaded network,
showing that it imposes the desired notion of fairness. Finally, we validate the connection-level routing policy proposed in
Section VI-B in an example where uncoordinated control is known to reduce the stability condition.
5.1. Controlling the Number of Connections
We implemented a packet-level simulation of the control law (10) in the network simulator [28].We have two
users that download data from three servers, with the topologies and link capacities depicted in Figure 4. In order to
introduce an imbalance between users, routes have different round-trip times (RTTs). Each user then begins with a single
TCP connection per route, governed by TCP/Newreno. With this choice, the congestion price pr on route is the packet-loss
probability along that route, and this is measured by the users counting the number of retransmitted packets within
connections.

Figure 4: Controlling the Number of Connections

5.2. Fairness via Admission Control
In this case, we simulated the linear network topology of Figure 5, with single-path users that generate
connections according to a Poisson process as in Section V. The network applies admission control using the rule, with
utilities chosen from the fair family with =5 to approximate max-min fairness. Simulations are performed in, and
individual connections are again controlled by TCP/Newreno. The entry router is in charge of measuring loss probability
along the routes. This is done in this case by sniffing the connections, although other approaches such as Re-ECN could be
used.

Impact Factor (JCC): 6.8785

Index Copernicus Value (ICV): 3.0

A Novel Analysis for Stability and Performance of Network

Resource Allocation and Num Problem in Large Scale Networks

49

Figure 5: Fairness via Admission Control (Fully Overloaded Case)

Figure 6: Admission Control (When the Source is Below Its Share)

In the second situation, we changed the load of user 2 to p2= 4Mb/s (below its fair share), and it is therefore
saturated as in Problem 3. The resulting rates are shown in the second graph in Figure 5. Here, admission control is
protecting user 2 by allowing its share of 4 Mb/s and reallocating the remaining capacity as in Problem 3 to 1*= 3*=
10Mb/s, 2* = 4Mb/s.

CONCLUSIONS
In this paper, we analyzed resource allocation in networks from a connection-level perspective with the intention
to bridge the gap between classical NUM applied to congestion control and the user-centric perspective. New notions of
fairness appear as user utilities are evaluated on aggregates of traffic, which can model different interesting situations. We
showed how the control of the number of connections can be used to impose these new notions of fairness, and how the
users can cooperate in order to drive the network to a fair equilibrium. We showed how admission control and routing
based on typical congestion prices can be used to protect the network in overload and simultaneously impose fairness
between its users. Finally, we showed practical implementations of the mechanisms derived in our work, and simulations
www.tjprc.org

editor@tjprc.org

50

Sai Ashwin Parakkal, Prasanna Balasubramanian & T. Saravanan

based on these implementations show that the proposals accomplish their goals.

REFERENCES
1.

D. Feijer and F. Paganini, Stability of primal-dual gradient dynamics and applications to network optimization,
Automatica, vol. 46, no. 12, pp. 19741981, 2010.

2.

Ferragut, Resource allocation in networks from a connection level perspective, Ph.D. dissertation, Universidad
de la Repblica, Montevideo, Uruguay, 2011.

3.

Ferragut and F. Paganini, Achieving network stability and user fairness through admission control of TCP
connections, in Proc. Conf. Inf. Sci. Syst., 2008, pp. 11951200.

4.

R. Khalili, N. Gast, M. Popovic, U. Upadhyay, and J.-Y. L. Boudec, MPTCP is not pareto-optimal: Performance
issues and a possible solution, in Proc. ACM CoNEXT, Dec. 2012, pp. 112.

5.

X. Lin, N. Shroff, and R. Srikant, On the connection-level stability of congestion-controlled communication

networks, IEEE Trans. Inf. Theory, vol. 54, no. 5, pp. 23172338, May 2007.

6.

X. Lin and N. B. Shroff, Utility maximization for communication networks with multi-path routing, IEEE/ACM
Trans. Autom. Control, vol. 51, no. 5, pp. 766781, May 2006.

7.

L. Massouli, Structural properties of proportional fairness: Stability and insensitivity, Ann. Appl. Probab., vol.
17, no. 3, pp. 809839, 2007.

8.

T. Bonald and L. Massouli, Impact of fairness on internet performance, in Proc. ACM SIGMETRICS/Perform.,
2001, pp. 8291.

9.

S. Boyd and L. Vanderberghe, Convex Optimization. Cambridge, U.K.: Cambridge Univ. Press, 2004.

10. Briscoe, Flow rate fairness: Dismantling a religion, Comput.Commun. Rev., vol. 37, no. 2, pp. 6374, 2006.
11. M. Chen and A. Zakhor, Flow control over wireless network and application layer implementation, in Proc.
IEEE INFOCOM, Apr. 2006,pp. 112.
12. M. Chen and A. Zakhor, Rate control for streaming video over wireless, in Proc. IEEE INFOCOM, Mar. 2004,
vol. 2, pp. 11811190.
13. M. Chiang, S. H. Low, A. R. Calderbank, and J. C. Doyle, Layering as optimization decomposition: A
mathematical theory of network architectures, Proc. IEEE, vol. 95, no. 1, pp. 255312, Jan. 2007.
14. G. de Veciana, T.-J. Lee, and T. Konstantopoulos, Stability and performance analysis of networks supporting
services with rate control Could the internet be unstable?, in Proc. IEEE INFOCOM, 1999, pp. 802810.
15. D. Feijer and F. Paganini, Stability of primal-dual gradient dynamics and applications to network optimization,
Automatica, vol. 46, no. 12, pp. 19741981, 2010.

Impact Factor (JCC): 6.8785

Index Copernicus Value (ICV): 3.0