A Scheme for the Quantification of Congestion in Communication Services and Systems

The characterization of congestion in communication systems is a pre-requisite for the definition of mechanisms that can be used in congestion control, with the goal to minimize its effects on the performance of distributed applications. In order for those mechanisms to work properly, there is the need to quantify congestion of services and systems. This paper proposes a scheme that enables the quantification of the congestion of a particular communication service, the comparison of the congestion degree of two or more services, and the quantification of the global system congestion, based on the definition of a set of system QoS parameters, of their normal variation limits and of their degradation thresholds. The paper begins with the characterization of the congestion problem. Following, a congestion definition is presented and a congestion metric is proposed. The paper ends with an analysis of the quantification of congestion in communication systems when looked at as a set of several consecutive modules. 1. Causes and effects of congestion The congestion phenomenon is associated with all communication systems of variable geometry, that is communication systems where there is a dynamic number of communication players (acting as information sources and/or destinations), where traffic characteristics can change, or where the available communication resources are not constant. This is the case of the majority of modern communication systems, for which congestion control has become a major issue due to the need to support a variety of applications with and without stringent communication requirements, over a variety of transmission technologies. From a global communication system perspective, and as a first step to the problem characterization, it can be said that congestion occurs whenever the total amount of traffic that enters a communication system in a fixed time interval is greater than the communication system outgoing flow capacity in the direction of traffic destinations, in the same time interval. Congestion affects the quality of communicating applications leading, in extreme cases, to their termination and results in resource wasting, as communication resources must deal with the original traffic overload as well as with the traffic that results from information retransmissions generated by losses due to congestion. Under working conditions, one can identify three distinct operating zones for a communication system: the linear zone, the congestion zone and the collapse zone . In the linear zone the throughput increases linearly with the system load, and the transit delay remains practically constant. In this zone, the communication system is being used well below its maximum capacity and, thus, it can rapidly respond to the load variation without significant impact on the transit delay. From the applications point of view, communication systems should be engineered and tuned to work in this zone. Nevertheless, cost/benefit considerations prohibit this, leading to a dimensioning that implies a non-zero saturation probability, which corresponds to a non-zero probability of entering the congestion zone. In the congestion zone, the saturation of the transmission resources due to the load increase leads to the intensive use of the communication system buffering capacity, which results in an increase of the transit delay and in a practically constant liquid throughput. It is in this zone that resource utilization is maximized, at the cost of a significant increase of the transit delay. In case there are no mechanisms to prevent the load increase when the system is already in the congestion zone, the system will enter in the collapse zone, which is characterized by the blocking of the communication resources due to intensive retransmissions generated by the overflow of the system queues. In this zone the throughput decreases (approaching zero), and the transit delay can increase up to extremely high values. In this case, all of the communication system bandwidth is being used for retransmissions. From a macroscopic point of view, the optimum system working point is at the threshold between the linear and the congestion zones. At this point the ratio between the liquid throughput and the transit delay also known as the "power" of the communication system [1] [2] is maximized. This situation corresponds to a compromise between the maximization of resource usage (which is desirable from the communication system perspective) and the minimization of the transit delay (desirable from the communication services perspective). The congestion control functions should lead the system to this optimum working point, controlling the system in such a way that avoids the entrance in the congestion zone, and preventing it from entering the collapse zone. But keeping the optimum system working point does not guarantee the fulfillment of individual requirements of the services supported by the communication system. Each service may have its own requirements in terms of the compromise between the liquid throughput and the transit delay, which may be different from the compromise adopted at system level. In addition, each communication service may have specific needs determined by other communication parameters, that can hardly be met by the traffic characteristics that result from the system optimum working point. Finally, a communication system can present a globally congestionfree behavior and, when analyzing each of the active applications, show great asymmetries in resource usage. Thus, in addition to search and maintain the global optimum working point, congestion control functions should contribute to the fulfillment of individual communication services needs, and avoid excessive use of resources by some services when compared to others. 2. Congestion definition As mentioned above, a globally non-congested communication system may behave as a congested one for one or more of its users. Thus, there is the need for a congestion definition that is able to cope with congestion from a global as well as from an individual point of view. The following types of congestion definitions are commonly found in the literature: •“a communication system is congested if the transit delay is greater than X” [2] [3]; • “a communication system is congested if the effective throughput is less than Y” [4]; • a combination of the above definitions [5] [6]. These types of definitions are not precise, as they evaluate congestion on the basis of one or more of its effects (increase of transit delay and/or decrease of throughput) without taking into account the main cause of congestion: the load increase. In addition, the determination of the congestion thresholds X and Y is subjective and/or difficult, which results in the inability to quantify the congestion degree of the communication system. Another limitation lies in the fact that congestion is evaluated from a global perspective, without concern with individual applications. Thus, it is perfectly possible that in a congested system (according to one or more of the above presented definitions) certain applications may be able to perform within the throughput and delay limits that suite their needs. In [7] a congestion definition is proposed that overcomes some of the above mentioned limitations. According to this author, a communication system is under congestion (from a user point of view) if the system usefulness decreases due to an increase of system load. The concept of usefulness expresses the user preference for the communication resources, through a usefulness function. This definition evaluates congestion from the users point of view, and characterizes congestion on a cause basis (as opposed to an evaluation based on the effects of congestion). Nevertheless, it does not support the quantification of the congestion situation, as the usefulness functions are user-specific and, thus, usefulness values cannot be compared in order to measure relative congestion among users nor combined in order to measure the global congestion of the communication system. Normally, communication systems and applications are dimensioned in such a way that the traffic alterations caused by the physical and technical limitations are within the limits tolerated by the applications and do not prevent their normal functioning. When the traffic characteristics are modified in a degree that decreases the performance of a given distributed application, the application is said to be affected by congestion. Thus, the congestion phenomenon can be characterized, from a communication service user point of view, by the following definition: Definition 1 congestion of a communication system: a communication system is congested whenever the functioning of communication services is affected in a way that is perceptible to their users. This definition emphasizes the communication services users perspective and accommodates all the factors that may cause the refusal, interruption, or degradation of the communication services, in addition to the load increase factor. In fact, this factor independence is consistent with the user point of view, to whom service degradation is the only effect he/she is concerned with, regardless of the factors that cause it. The proposed definition has when compared to the previous definitions based on throughput variation and/or transit delay the advantage of characterizing congestion from a microscopic point of view for each of the supported services, and at each instant of time as opposed to a macroscopic characterization based on the global throughput and transit delay. Nevertheless, the proposed definition is subjective and needs to be complemented by a metric, in order to enable the quantification of the congestion of communication systems. 3. Congestion metric From a user perspective, a communication system without congestion is characterized by the ability to provide the quality of service (QoS) required by the active applications. Quality of service can be objectively defined by a set of operating parameters or, implicitly, by a set of values that are considered "normal" when the application is active. On networks that are based on the resource reservation paradigm normally operating in the connection-mode it is always possible to objectively establish a set of quality of service parameters, because these are the parameters that are necessary for the resource reservation at the time of connection establishment. On networks that are based on the best effort paradigm normally operating in the connectionless-mode there is no need to explicitly define the QoS parameters; nevertheless, they can be implicitly deducted from the evaluation of the behavior of applications. In either case best effort or resource reservation it is always possible to determine a set of parameters the QoS parameters that are responsible for the characterization of the quality of service of the supported applications. Let PQoS , defined in Expression 1, be the set of all of the QoS parameters supported by a given communication system. The description of the physical significance of each parameter and the identification of its respective units is considered to be associated with the definition of the PQoS set. PQoS = q1,q2, q3 , .. .,qn { } (1) For each of the supported services si , and at discrete time instants tk (or continuously in time), it is possible to measure or compute a set of values (one for each parameter belonging to the PQoS set) that can be stored in a vector, as shown in Expression 2. VQoS(si )t k = v1 v2 v3 . .. vn [ ] (2) The specification of the quality of service necessary for each user application may be done, for each parameter, by the specification of an interval in which the parameter values imply no QoS degradation, and the specification of one lower threshold and one upper threshold beyond which the quality of service is unacceptable. The set of interval limits and degradation thresholds, specified for a given service si, may take the form of the matrix presented in Expression 3 the QoS matrix, MQoS in which mj and Mj are, respectively, the minimum and maximum values that parameter qj may take without QoS degradation, and lmj and lMj are the thresholds that subtracted from mj and added to Mj, respectively, define two operating zones with degraded but still acceptable quality of service. MQoS(si) = m1 lm1 M1 lM1 m2 lm 2 M2 lM2 m3 lm3 M3 lM3 : : : :