Internet Resource Pricing Models, Mechanisms, and Methods

Too many packets will incur network performance degradation, which is called congestion [1]. Congestion is caused by unbalanced resource and traffic distribution, and thus will not be automatically eliminated with the increase of network capacity. In packet switched network, the selfish nature of users makes this happen. As shown by Hardin [2], “tragedy of commons” occurs when many individuals share public resources and each holds a selfish objective, which means the loss they bring to others is larger than their own improved benefits. So, if the network is used as public goods, there always exists the possibility that the overall personal excessive usage will cause system performance decline and thus the congestion problem.
In recent years, high bandwidth, low latency, low jitter and other higher QoS applications are getting increasingly popular. Thus the surges of network traffic makes network congestion more frequent and serious. Accordingly, the novel content distribution technologies and mechanisms to ensure network QoS are constantly proposed and improved. For the former, commonly, a new layer of network architecture, the application layer network (Overlay Network [3]) is added in the existing Internet to realize the corresponding transmission and QoS control, such as P2P (Peer-to-Peer) [4] and CDN (Content Distribution Networks) [5]. For the latter, mechanisms are developed to work at all levels of QoS control, such as transport layer and network layer concerning network service structures. In short, they both serve network resource management and congestion control.
However, on the one hand, network traffic surges and keeps increasing. As Valancius [7] shown in Fig. 1, videos and P2P traffic occupy a large part of network resources and will become even more in the coming years. On the other hand, different application layer networks have their own selfish traffic demands and QoS control mechanisms. This makes network management and maintainence increasingly difficult [6]. As an earlier best-effort network service type, Internet Service Providers (ISPs) often meet the increasingly high QoS requirements by upgrading network infrastructure or increasing network capacity. However, in the long run, short-term investments usually bring high cost and fail to satisfy the fast-growing network resource requirement, which is against the healthy network development. Therefore, QoS control technologies need to be introduced in best-effort network. From the perspective of improving network resource usage and management, network designers and ISPs usually passively conduct QoS control based on the existing network traffic, such as congestion control [8]-[10], and traffic engineering [11]. But these often complicate network protocol design and implementation. Proactively setting QoS levels of flows for simple QoS control (priority-based QoS mechanism [13][14]) and designing network architecture to ensure QoS (such as IntServ [15] and DiffServ [16]) have also been studied extensively. But due to some technological limitations and lack of incentives, they have not been implemented throughout the network.

From the above discussion, we note that design incentives at economical level to encourage ISPs in improving network performance and directing users to use the resources rationally, will be of great significance in effective network resource management and distribution [39]. Such methods are based on the utility optimization theory in economics, which affects users’ demand and belongs to active resource management mechanism. Simply speaking, ISPs can effectively influence users’ demands and network resource usage by choosing rational pricing strategies, thus prompting efficient network usage and ensuring network performance. Particularly, as an important auxiliary aspect of technological progress (economic incentives [22]), pricing mechanism studies suited to service type development are also important. Therefore, a complete picture of network pricing should include three aspects: basic pricing models, mechanisms to ensure pricing implementation, and methods determining optimal pricing levels.
Specifically, first of all, pricing models decide which factors to charge, or how to evaluate network operating and maintaining costs. Mason and Varian [18][19] analyzed the major fee component from users’ cost point of view. This includes: a fixed fee to provide basic service structure costs such as leased lines, routing equipment maintenance, and human resource utilities; marginal costs of access; network expansion costs; marginal costs of sending data packets into the congested network; and social costs that cause negative impact on other users. The authors believe a good price should reflect these costs. So, we introduce three basic pricing models concerning these costs: flat pricing [18], usage pricing[21][22][25] and congestion pricing [18][29]-[37].
As applications are simple and resources are sufficient at the beginning of the Internet, it is convenient to charge users using a static flat pricing model, where users have the same usage-irrelative fixed fees with equal access rates. The advantages are that complex audit and statistics are unnecessary, and thus facilitates network users. So, they increasingly enrich network contents. However, too many contents eventually causes network resources lacking.And the defects of flat pricing gradually emerge. For the system, due to lacking of incentives for efficient network resource usage [20] (a lot of bandwidth are wasted by non-critical applications), the overall network performance degrades. For users, the experience deteriorates and the fairness cannot be guaranteed. Obviously, flat pricing is no longer applicable. Thus, a more effective resource pricing model “usage-based pricing” was proposed [21]. It pointed out that if the charge is related with usage, fair and efficient use of resources will be promoted to some extent. However, with a further increase in network traffic, the aggravated congestion makes the related pricing a hot research area, resulting in a relatively dynamic pricing model “congestion pricing” [18][19] which are studied extensively. Besides, these three pricing models can be used in any combination since they reflect different cost components.
As for pricing mechanisms, they mainly aim to address the matching problem between network service types and pricing models. Namely, for different types of network services, we need to select and design suitable pricing models. Good pricing mechanism can set rational price structures for users and ensure pricing implementation with an acceptable technical complexity measure [12]. Generally, in best-effort network, ISPs always adjust the basic pricing model to promote the rational use of resources based on their network capacity, where no additional QoS control mechanisms are conducted. Odlyzko’s PMP (Paris Metro Pricing [55]) pricing aims to achieve QoS differentiation and thus enhances efficiency through dividing network into several subnets in best-effort network. However, with the increasing emphasis on applications’ QoS and network resource usage efficiency, network designers and ISPs both tend to serve different data streams with different QoS and price levels. Simple priority-based pricing was first proposed by Cocchi et al. [13] [14]. The authors suggested to implement prioritized service using priority field in IP packets, and thus they can conduct service layering and corresponding pricing. Similar thoughts can be found in [42]. With progressive development of various network service types, QoS guaranteed network architectures (such as IntServ and DiffServ) are gradually studied in recent years, followed by corresponding pricing models. QoS based network resource pricing mechanisms are thus formulated [43]-[51][56]-[60]. We discuss pricing models suitable to various service types in Section 3.
For the last aspect of pricing strategy, pricing methods applicable to pricing model/mechanism are still an important research aspect. It mainly determines how to set a reasonable price level. An ideal pricing method should be able to set price levels that can control resource usages so as to achieve its pricing objectives while achieving efficiency. Determining prices is usually based on relevant fields of pricing and utility optimization in economic theories under specific market environments. Such work is often based on different market structures (such as monopoly and competitive network) and network service mechanisms (such as best-effort and QoS guaranteed service network). After studying each entity’s utility, different theory models are used to describe their interactive optimization processes. The theoretical bases are mainly optimization theory and game theory. Thus there are two major research lines: (1) Studying pricing based on system optimization (Network Utility Maximization, NUM [30][31]) always lies in optimization theory [74]; (2) Studying pricing based on strategic optimizations of ISPs and users. That is, when analyzing each player’s decision making, one should take into account effects from strategic behaviors of other players. This work is mainly based on two major theoretical branches of game theory: non-cooperative game theory [75][84] (related models such as in [77]-[82]), and cooperative game theory [83]-[85] (related models such as in [87][91][92]).

As shown in Fig.2, the remainder of the paper presents a detailed survey on Internet pricing development. In Section 2, we present three main pricing models proposed in earlier years. Then, integrated with pricing models, we introduce pricing mechanisms based on two types of services in Section 3. In Section 4, we introduce price level setting methods based on two classes of optimizations, including system optimization and entities’ strategic optimizations in different network marketing environments, which can economically incentivize technology development. We classify and compare typical pricing strategies in Section 5 based on different pricing models, serving mechanisms and pricing methods involved. Finally, in Section 6, we conclude the paper, predict the reasonable pricing strategies for new applications and network services, and point out several future research topics.

At the early stages of Internet, users use a small quantity of network resources. Thus ISPs aim to attract a large number of users and occupy the market. They generally adopt unified price (or flat fee [18]) to charge users based on access costs, which means in a certain period of time, the users with the same access speed will be charged at the same price. This is especially common in broadband access market.
The advantages are as follows. For ISPs, flat pricing is popular, since it is easy to implement and there is no need for complex statistical systems. And for users, the charges can be predicted. However, the more usage, the more obvious drawbacks. On the one hand, due to lack of effective interactions between users and ISPs, users have no incentives or ideas about adapting their usage patterns, making network resources over requested or used. On the other hand, ISPs do not count individuals’ resource consumptions and treat equally to users with the same access rate level. This means the overall cost is equally shared by users with different consumptions and thus fairness is hard to guarantee. Meanwhile, as there is no difference in charging users, ISPs lack impetus for upgrading infrastructure or improving QoS, which is not conductive to the progress of network technology and makes system performance degrade.
As to the fairness, Edell and Varaiya studied users’ reactions on flat pricing through Internet Demand Experiment project (INDEX [20]). They concluded that light-load users compensate the heavy-load ones under flat pricing, which will cause resource waste too. The authors assumed unit usage cost is charged by $c$, and users request $D(c)$ unit resource according to demand curve. As using flat pricing model, the marginal usage cost for users is 0, which makes the demand changed from $D(c)$ to $D(0)$. Estimated by users’ practical utilities, the usage over $D(c)$ will cause ${\int}_{0}^{c}[D(c)-D(0)]dp$ value loss to users, as the shade shown in Fig. 3 [20]. In addition, if the flat fee $C$ is charged based on average usage amount, then $C=c\times x_{f}(av)=c\times D(0)$. All users’ payments are shown as the rectangle area in Fig. 3 [20]. Clearly, the light-load users’ payment is more than their gain, while the heavy-load users are on the contrary. This indicates that the former compensates the latter when they share resource costs on average.
As discussed above, in best-effort network without additional QoS mechanism, flat pricing model is unable to achieve optimized resource allocation alone. And due to fewer ISPs, the marketization is not obvious, which worsen the situation that ISPs lack incentives to improve network performance. With the development of network applications and the increasingly complex Internet marketing environment, the model will no longer apply. But as one of the referential pricing factors, access charge can be used as a basic guarantee for recovering the fixed costs.

As the usage and fixed costs have been distinguished and studied separately, usage-based pricing models come into being. Currence et al. [22] thought usage-based pricing can reflect actual use of network resources and is derived from traditional flat pricing. Simple usage-based pricing uses the amount of upload and download traffic to charge.
In practice, China Education and Research NETwork (CERNET) uses full-rate accounting charges for international traffic [24]. In addition to such direct traffic statistics, ISPs in general can use statistical sampling methods to estimate usage, such as the $95^{th}$ percentile pricing which is used as an industry standard. This is in accordance with usage-based pricing, and the peak flow within 5% of the time (36 hours per month) is free of charge. Many ISP, such as MCI WorldCom and Level (3) Communications, have such peak flow rate based charging standards [22].
Usage-based pricing is analyzed and studied by a lot of researchers at early stages of the Internet [12][21]-[25]. The common point is that in general they used supply-demand balance models in economics to describe the interactions between users and ISPs. Edell and Varaiya [20] showed in their experiments that users are highly sensitive to pricing models and price levels. Usage-based charging, can not only enhance usage efficiency of network resource, but also play an important role in congestion control and fairness guarantee among users. Edell et al. [21] implemented a usage pricing system and gave experiments illustrating that dynamic usage pricing can prevent congestion and improve the average network performance. Courcoubetis et al. [27] proposed intelligent agents to decide network usage, based on network conditions and users’ payment willingness. This simplifies users’ utility optimization process.
After analyzing the features of flat and usage pricing models, Altmann and Chu [23] proposed a hybrid pricing model that combines two. In this novel model, users enjoy basic services at a basic flat rate, while higher bandwidth demands will be charged by usage. The experimental data analysis indicates that such pricing model can improve network performance and increase ISP revenue. Obviously, such pricing concerning fixed and usage cost will benefit all the participants.
Recently, with the continuous development of high-bandwidth required applications and P2P content distribution technologies, the overall users’ bandwidth demands increase dramatically. Consequently, increasingly differentiated usage patterns make the fairness problem even more serious, which indicates charging heavy-load users according to usage is more reasonable [26]. However, in terms of P2P applications’ providers who encourage users to participate in content sharing, such charging scheme will go contrary to their goals. So, more complicated interactions between P2P application providers and ISPs are to be carefully studied. In addition, other problems still need to be addressed, such as the privacy issues in processing audit and statistics [22] and the charging problem caused by users’ non-expected traffic (such as ads and spams).

The pricing models mentioned above cannot reflect individual traffic’s impact on network, such as packet loss and delay. An intuitive understanding is that, too many concurrent network users will easily degrade network performance. For those who have accessed in network, the higher system load, the higher possibility of congestion. This also means that more external cost will be caused by users [18].
Researchers expect pricing can constrain this negative external effect which is also called social cost. And the corresponding pricing is named congestion pricing [18]. Congestion pricing dynamically sets price that can reflect approximate real-time network resource usage and represent current social cost. Thus it can encourage users to adjust traffic demand which may avoid excessive resource usage. Therefore, congestion can be relieved or eliminated [29]-[37].
However, measuring such social cost is not trivial. It cannot be directly calculated or measured as fixed or usage cost but need to detect users’ perceived value of resources. In general network performance optimization articles, congestion cost is described by delay in M/M/1 queuing system [79]. In Mason and Varian’s smart market [18] pricing mechanism, an auction based pricing method was proposed to measure and price such social cost. As shown in Fig. 4, the steps are as follows: (1) Users fill in bid fields for each packet on behalf of their willingness to pay for the packet transmission, e.g., P1(3) represents user 1’s willingness to pay for its packet is 3; P2(30) means user 2 is willing to pay 30 for its packet; P3(20) means user 3 is willing to pay 20. (2) The routing node (auction point) receives the packets and sorts the packets according to the bid values. (3) Checking available bandwidth, the routing node sets marginal bid value as the market clearing price or threshold price, and decides which packets to be transmitted (or discarded). In this example, we see that if the node can only process two packets, the packets from user 2 and user 3 will be transmitted at the price of 20 for each packet, otherwise discarded.

This section describes basic pricing models based on cost analysis in traditional best-effort network. They are gradually proposed and thoroughly studied along with the increase of network resource usage. Obviously, with the increasing importance of pricing in effective network resource management, pricing models will consider more factors and be more complex. From performance optimization perspective, this section describes pricing models with nearly different functions. In a flat pricing model, the fee is generally constant in a long period of time and is used to recover the fixed cost. Usage-based fee is charged to recover usage cost. It can be adjusted to reflect network congestion and thus plays a role in congestion control. Congestion pricing is proposed to measure and charge for congestion. It is a kind of dynamic pricing where price is dynamically adjusted to congestion.
In fact, these three pricing models are not orthogonal, which means although they reflect different pricing factors, their functions can be overlapped to some extent. For example, “two part tariff” [19] was proposed as a combination of flat pricing and usage-based pricing. It can reduce congestion to some extent. In addition, congestion price mainly reflects the marginal cost of lacked resources. It can also be interpreted as the potential benefit increase of network users if there is one more resource unit. Therefore, congestion price is closely related to the timely network resources usage.

In best-effort network, as ISPs generally do not implement additional QoS control mechanisms, there is nearly no QoS difference. Thus, ISPs adjust basic pricing models to affect resource usage while optimizing economic benefit. Pricing is always done at network edge, known as edge pricing [38][39]. It means that users’ fees are calculated by the access network but not directly concerned with intermediate networks along the whole transmission path.
Supporters hold the following beliefs. On the one hand, Internet users located in different autonomous system are often managed and charged by local ISPs. Thus it is more realistic to charge users at the access side. On the other hand, as a best-effort service, ISPs provide no QoS guarantee to whatever traffic traversed through their networks. So pricing at network edge is more reasonable [38][39].
The basic pricing models suitable to edge pricing include flat pricing and usage pricing. For congestion pricing, because congestion could occur in any link along the transmission path, the price should be set according to path usage status. Thus it is not applicable to edge pricing. Moreover, for data packets, there may be multiple paths to select. But routing or path is not decided by users. So it is unfair to charge them for the path they use [38]. However, Shenker et al. [38] pointed out that edge pricing can still refer to approximate congestion and users’ expected paths.
Clark [39] further discussed localization method for non-local accounting and pricing, such as setting price for multicast users and pricing for receiver-paid applications. The method is based on resource reservation protocol (see Section 3.2.2), where the sender first chooses how much it will pay or what portion of cost to share with receivers. In [40][41], Clark suggested edge pricing could use estimated traffic instead of actual usage to charge users. And receivers can also state their willingness to pay. ISPs exchange traffic and revenue through agreements. Later, when bandwidth management devices [42] are added in the DiffServ architecture (see Section 3.2.3), the relatively dynamic edge pricing based on expectations or estimations is also being studied [43]. However, obviously, the edge pricing lacks influence on congestion control. Yuksel and Kalyanaraman [44] proposed a distributed dynamic pricing that is congestion sensitive and whose sensitivity and complexity are ranged between those of “smart market” and edge pricing.

Facing unachieved QoS differentiation and corresponding low network resource usage efficiency, a lot of work has committed to study of differentiated services so as to enhance efficiency. Simple priority-based service and corresponding pricing were first introduced by Cocchi et al. [13][14], which revealed the relationship between QoS differentiation and resource usage efficiency. They proposed to add priority field in IP packet and achieve QoS through priority-based queuing and scheduling. The corresponding service pricing is thus being wildly studied [55]-[60].
Then, with the progressive development of various network service types, to achieve QoS guarantee, various in-depth studies were conducted regarding network architecture based on resource reservation [15] and flow aggregation [16]. Also, related pricing models are studied and integrated into such QoS-enabled pricing mechanisms [45]-[51].

Based on two types of network services considering QoS or not (best-effort service and QoS guaranteed service), we introduce two kinds of pricing mechanisms in this section. For the former, edge pricing is a relatively suitable implementation, PMP has also been proposed as a variation. And for the latter, we introduce the pricing mechanisms proposed mainly within the scope of pricing, which serve two service architectures named IntServ and DiffServ.
For best-effort service, Shenker et al. believed that if edge pricing uses expected congestion information, it can also achieve a certain degree of congestion control. Also, one can distinguish access bandwidths to achieve some kind of prioritized services. But both of them cannot assure the usage efficiency of resources and guarantee QoS or network performance.
For QoS-based service pricing, as QoS is differentiated by packet processing based on service classification or resource reservation, which often needs support from devices or networks on the entire transmission path, and thus is more complex than the former pricing mechanism. Especially for IntServ pricing, as QoS is guaranteed based on resource reservations where the service mechanism itself is complex, the corresponding ricing process can be even difficult with higher complexity. However, when it comes to service differentiation or DiffServ, from the perspective of efficiency, to a certain extent, we can conclude that it facilitates the efficient use of resources (high QoS requiring packets are prioritized processed) and ensures fairness among users (i.e., which service class or agreement is chosen by users), though there is no assured QoS guarantee. Indeed, combining IntServ (in edge network) with DiffServ (in core network) to provide differentiated services can have low complexity and improve efficiency with a certain degree of QoS guarantee.

From an economic point of view, efficient market means total social surplus or the sum of service providers’ surplus and users’ surplus is maximized [61], which equals to the difference maximization between the value of resources to users and the cost of providers. Under different market environments, different conclusions can be drawn. We mainly introduce system utility (social surplus) optimization oriented pricing method for a single network based on the optimization theory. The system is consisted of users with different utility functions and a network with resource constraints [29]. In fact, this research line has a tremendous influence on communication networks. It prompted an in-depth understanding of network architecture and a guided protocol design for more efficient network resource usages.
Kelly [30] proposed the concept of Network Utility Maximization (NUM) which is the initial work of Internet system optimization. In his work of pricing and resource allocation, the main object is to find the price that can make the total resource demand and supply in equilibrium. According to market pricing theory in [56], if a system is in demand- supply equilibrium, the system utility or social surplus will be maximized. NUM framework can be described by three optimization problems. The system optimization is a radical problem which can be first modeled as: $\hbox{maxmize}\ \underset{s}{\sum}U_{s}(x_{s})$ (the service provider’s cost is ignored), where $x_{s}$ denotes the traffic rate and $U_{s}$ denotes the value or utility of the traffic to the corresponding user. The constraints are: (1) $Hy=x$, where $H_{s\times r}$ denotes the source-destination pair $i\in\{1,2,\cdots,s\}$ is served by path $j\in\{1,2,\cdots,r\}$, and vector $y=\left\{y_{1},y_{2},\cdots,y_{r}\right\}^{T}$ denotes the resources distributed to all source-destination pairs on each feasible path. This constraint means the whole distributed resources equal to $x_{s}$ for any user; (2) $Ay\leq C$, where $A$ is a 0-1 matrix telling whether the distributed resource is on the link, and the constraint means the sum of all the distributed resources will be no more than link capacity $C$; (3) $x,y\geq 0$. The above can be rewritten as the whole problem (A):

Within a single ISP network, system equilibrium based on supply-demand relationship can ensure optimal pricing which maximizes system utility in Section 4.1. This type of equilibrium is achieved through pricing where the ISP and users indirectly interact with each other. However, in real network, there are three types of relationships: ISP-ISP, ISP-users, and user-user. Most of them are modeled by considering their direct interactive effects.
Game theory studies how individual decision is made considering others’ actions. And it also predicts whether there exists an equilibrium under such strategic behaviors. The utilities in this type of model are directly affected by the other participants’ strategies or preferences. Thus when studying the direct effects among network participates’ behaviors, game theory is always used as a basic theory. Based on whether a binding agreement can be formed, games are divided into non-cooperative games [75][86] and cooperative games [83]-[85].

We classified and summarized typical pricing methods of network resources based on two main research lines. The main points are as follows:
(1) System optimization model mainly based on the NUM framework. Considering network traffic characteristics, it can be divided into: i) Optimization model for elastic flow system; ii) Optimization model for hybrid system where inelastic and elastic flows coexist.
This class of work is usually based on supply-demand relationship. It aims to find the optimal price and rate allocation with balanced supply and demand where the maximal system efficiency is achieved. As inelastic traffics (such as real-time video and voice flows) emerge and largely increase, system optimization for hybrid network has drawn more and more attentions. Compared with elastic flow system optimization which has unique optimal solution shown by convex optimization theory, inelastic flows are always described by $S$-type utility function. Thus it turns the system problem into a complex non-convex optimization problem. Therefore, price-based access control is mainly introduced here to assist resource distribution. It generally includes two methods: users’ self-regulation [65] based on their own utilities and the network conducted access control based on network efficiency [67]. Hande et al. [68] considered elastic flow protection in hybrid system. They believed that the elastic traffics are less competitive than the inelastic flows.
In short, optimization-based modeling for hybrid system has high complexity, and especially hard to solve in real systems. Also, as on a network transmission path, access control policies of each link may be different, there lacks well-designed distributed decision-making mechanisms which can ensure system convergence to the global optimal rate allocation.
(2) Strategic optimization model based on game theory. Based on two major branches of game theory: non-cooperative game and cooperative game, we introduce some typical corresponding models used in network resource pricing.
As modeling for strategic interactions, non-cooperative game model mainly discusses NE point and its characteristics. Cooperative game model we introduced here emphasizes the fairness criteria in sharing. The point is that the solution of the former may not be Pareto optimal, however, the latter sometimes needs constraints from a third party to ensure cooperation.
Comparing the above system optimization model with the non-cooperative game model, it is obvious that different model ideas always need support from different theories. In NUM framework based system optimization, the equilibrium is achieved by indirect interactions between the network and users based on price. And in this process, ISP dynamically controls the system through pricing mechanism to help reach an optimal equilibrium. Non-cooperative game model directly analyzes pricing problem based on strategic behaviors of all participants, which can quickly determine whether the system has NE point or not. However, it is possible that the equilibrium point exists but is not achievable. Then, the uniqueness and stability of the existed NE will also be discussed mainly using optimization theory. In fact, if models have equal essential meanings for key parameters (e.g., revenue and cost), the results based on different basic theories (NUM or non-cooperative game) are nearly the same.