1.1 Users' Involvement

Users of the Internet make choices based on their individual value of information and the cost of waiting for responses to the various information requests. Their selections of Internet server to use, time to submit request, the size of the request, and potential abortion of submitted requests, have an indirect influence on their individual allocation levels as well as externality effects on concurrent users sharing the same resources. As such, the system is slowly self-regulating at the user level, complemented with faster regulation at lower layers, such as roundtrip-time window adjustments at the transport layer.

With recent proposals to charge users for the network resources they consume, an alternate opportunity emerges for control of IP allocation that can individualize allocations at the user level. Briefly stated, the proposals promote a slowly-changing fee per IP packet for production cost recovery and a rapidly-changing congestion fee per IP packet for dynamic balancing of usage value and waiting cost. High uncertainty as to the magnitude and variation of the IP usage costs motivate a need for users to control expenditures.

The design of an Expenditure Controller Interface (ECI) will be proposed, via which the user can specify a real-time upper bound on his expenditures. His decisions will be based on quality perceived and expenditure feedback. Since the bound regulates the flow of IP packets from the user's active application, the ECI and the user form a response to the pricing process and contribute to load regulation on the current IP path.

1.2 Computational Markets

In general, the coupling of congestion fees and users with appropriately designed ECIs defines a real-time market for resources whose potential use for computer networks was noted in [3, 9]. Current work has described the potentially very efficient Smart Market trading mechanism where users submit sealed bids per packet and the service center decides job schedule and optimal price. This has been examined for energy networks in [25] and IP networks in [20]. Related mechanisms for access to distributed and parallel CPU resources were demonstrated with favorable experimental results in [31]. Economic models have been developed under the assumption that the price coordinator has estimates on users' valuation on waiting time [27]. Dynamic pricing was shown, via simulation, to improve on net consumer surplus produced in an Internet [13] and on cell drop rates in an ATM network [24]. Typical for these papers is an emphasis on the market mechanism rather than user involvement, and the assumption of "smart agents" at the user site, dealing on behalf of the user [13].

1.3 Human Process Control

Direct user involvement for real-time price-quality control in digital telephony was proposed in [7]. In an environment where price and quality change frequently, and telephones have speech compression equipment that removes quiet periods, the user would regulate the price-quality relationship by talking more or less. The costs could be controlled by shifting speech activity to periods where talking is cheaper. His real-time decisions would be motivated by real-time feedback; costs of talking would be displayed and speech quality would be detected in the form of the received earpiece speech quality.

Hume and Lewis studied the use of human control of multiple file transfer scheduling over a congestible and dynamically priced network [14]. In their evaluation, process control problems appear to be advantageously automated when the task involves highly procedural observation, storage and computation with quickly changing numerical input data, whereas the selective and inductive behavior so characteristic of humans appears more suited for high entropy tasks that are hard to characterize. Their work demonstrates the use of ecological interface design theory to significantly improve the operator's ability to schedule file transfers within the deadline and monetary budget.

1.4 Quality Control

Drawing from this, it appears that control of expenditure in real-time falls within the domain of control problems better suited for automation. Expenditure is the price times usage integral over time, where price changes very rapidly and the usage is generally unobservable by the user. The ECI will automatically ensure that expected expenditure rate does not go beyond a user-specified upper bound. Our design leaves the setting of the upper bound for the user. Optimal setting of this bound depends on the individual quality preferences of the user and trends in the price-quality relationship. Users perceive quality differently depending on the type of application they are running.

Recent trends have favored an approach where the application specifies its service requirement to the network using the service descriptors available from the network interface. The service requirement asked for by a request-response type application (e.g. telnet, ftp and www navigators) is based on studies of user preferences for response times to various types of information requests. For continuous media, the service requirement is based on user's tolerance to noise in the received information (see [3] for details). Pre-programming service requirements may be a technically simple solution to provide the average user with the average preferred service quality, but allows for suboptimal over- or underprovisioning of resources to nonconforming users.

The ECI proposed in this paper requires no user quality specifications to performs its role. All that is needed is the upper limit on expenditures which in turn bounds IP resource usage. The sufficiency of this approach lies in that response time can be reduced by increasing the processing resource allocation to a user [28], and noise can be reduced by allocating either more buffers or processing resources to the user [18]. Since the user at any time can adapt his expenditure rate control policy to his current private quality preferences and budgets, the ECI is an application independent quality control methodology.

1.5 Accounting and Billing

Computational markets rely on usage prices to be computed by the service center, i.e. an IP packet switch, and for endpoints, i.e. users, to respond. Sufficient user incentives to regulate in real-time can be preserved through (a) reliable billing, which requires an accounting mechanism such as in [4, 23], and (b) virtually instantaneous expenditure feedback, which requires the ECI to know the current packet price and billing rule.

A standard IP billing format is not defined yet. Ad hoc solutions for cost recovery are used in Chile [1] and New Zealand [5]. Based on the accounting architecture, the local Internet provider can bill for all incoming and/or outgoing packets, but not identify the originator of an end-to-end call, which eliminates billing of the calling party. Higher-level distribution mechanisms for IP usage costs are likely to emerge. A possible scenario is that the user is billed for his outgoing traffic (in line with [23]), and that the price he pays per packet is billed only once.

For congestion control purposes, individual IP switches may exercise intra-AD backpressure price control. In this case the user will pay the usage price of the closest IP switch, and the price may reflect both congestion cost as well as forwarding costs to neighboring IP switches. The purpose of the ECI is to facilitate efficient load control for high frequency pricing mechanisms, so it is largely irrelevant to this paper where the price is computed as long as the ECI knows the rule with which it will be billed.

2.1 Measured Rate Services

Indications of a "discomfort" with measured rate services has been voiced on various technical Internet mailing lists and in other media (see e.g. [1, 33]). Users seem inclined to prefer a pricing structure where the organization pays a capacity-dependent monthly fee for access without any fee for usage (see [21] for a review of the current practice). This is somewhat reminiscent of the debate over mandatory local measured telephone service that took place in the early to mid 1980's.

Wenders [32, Chapter 3] argues that pricing in telecommunications must be separated into usage and access pricing components. [1] More germane to this discussion, Wenders [32, Chapter 5] argues for mandatory local measured service. Under this pricing regime, access prices are lowered, allowing for an increased number of users on the network, while the usage charges suppress usage. This has the effect of reducing the usage costs (because of suppressed traffic) even when the costs of the measuring equipment is included. While Wenders considers peak-load pricing, not dynamic pricing as we do, and while a telephone call is a more tangible commodity to the end user than a packet as part of an application is, his analysis is nonetheless instructive.

2.2 Analogies

When a user specifies an upper expenditure limit via the ECI he constrains the flow of IP packets from his application, without actually knowing by how much. He can only observe quality effects from the action. In a useful analogy, the accelerator pedal serves as the user tool for speed control in a car. Unaware of the actual gas consumption rate, the user has to learn the short run relationship between action and quality and, in the longer run, the effect between driving habit and monetary outlays. We expect this to hold for users of the ECI. Over repeated trials, users will learn how to operate effectively in their self-interest.

Simplicity of the ECI is deliberate for the initial design, although "smart agents" can reduce the mental burden on the user, and have been called for in general computational economies [13, 26, 31]. We choose to fit the control problem within what we think are current ranges of human capability and trust that automation of subtasks will be pursued in future works.

2.3 The ECI Model

A user will select to submit an information request if his expected net benefit is positive. Let v be the information value of the information item when successfully received by the user, C(T) be the congestion cost from having to wait T time units for the response, and P be the IP expenditure. Since neither T nor P is known perfectly, ahead of the submission, the user must deal with expectations, i.e. submit if v - C() - ≥ 0, and form those expectations through learning.

A congestion priced resource is assumed to announce a new packet price at fixed intervals in time. In response to this price process, and while awaiting a response to a submitted request, the user controls three variables: A budget unit m> 0, a current budget multiplier u ∊ (0, 1,..., K) and a budget period length k> 0. The ECI ensures that IP packets are not submitted when > , where is predicted period expenditure and = mu is period expenditure bound set by the user. While all three elements can be specified by the user at any time, it is expected that the user rarely will specify k, occasionally will specify m, and will use u as the real-time control tool. Perhaps the simplest control interface is binary, with K = 1. The user either allows for expenditure (u = 1 and = um = m) or not (u = 0).

2.4 User Control in The Smart Market

The discussion so far assumed that the ECI is periodically informed of the new packet price (delayed by ∆). In the Smart Market the price is decided at each service time and cannot be publicized ahead of service, so the ECI needs tailoring.

Each IP packet submitted for service has a bid in its header. Let the set A(t) contain the packets awaiting service at time t at the IP switch. The next served packet is the one with the highest bid, for which the originating user will be charged a price equal to the lowest bid in A(t), which is at most equal to the bid of the packet served. The Smart Market thus guarantees an upper bound on the packet charge. [2]

A user can control the upper expenditure bound to for N IP packets by setting = /N. Since N is unobservable, unless the application designer provides an estimate, the tailored ECI for a Smart Market must provide for (a) a real-time controllable through which the user can manipulate quality and (b) expenditure feedback. While this is in line with the original Smart Market proposal, its proponents suggest reverse packets to carry the realized prices [20]. However, not all flows see reverse traffic, so the expenditure feedback must constitute an upper bound. The ECI knows how many packets were submitted at which bids and, due to the price guarantee, knows this upper bound. Long run differences between bidding levels and packet charges, as materialized in periodic billing information, are likely to influence the behavior of the user, e.g. by him bidding higher for better quality, or by using more resources.

2.5 Other Technologies

User involvement as discussed here can be used for real-time purchase of goods in connection oriented networks. When rate is dynamically priced [24], the user can regulate his rate requirement in real-time. More complex, when both rate and buffers are priced [19] the user may control both or some composite thereof. Again, both rate and buffer allocation form incomprehensible input goods, so the user has to learn the price-quality relationship and therefore relies on explicit expenditure feedback from the ECI and implicit quality feedback through the application.

2.6 Implementation

An implementation of the ECI requires modifications at both the application and transport level. The following figure illustrates a scenario with TCP as the transport layer.

The user perceives quality q from the application and possibly δ, the completion ratio of a bulk transfer. He applies control c towards the ECI, being (k, m, u) for the general expenditure rate controller or for the Smart Market expenditure per-request controller. His decisions are based on expenditure feedback which in the general case is xp or an upper bound x for the Smart Market, where x is actual usage as reported by TCP. The ECI may supply via a textual or a graphical information display and receive c via keyboard entries or window selections.

While TCP currently defers IP submissions when the congestion window or the receiver window is empty, it must now perform an additional test on an ECI-controlled variable . ECI updates every k time units as = /pwhere = um and p is the prevailing network price. If surplus carry-over is allowed, the ECI sets = ( + a)/p where a is the surplus of the previous period. [3]

In the additional test, TCP will submit an IP packet if > 0 and, if so, decrement by one. In the Smart Market, the user does not need to constrain usage rate so = ∞, but must be known by TCP for insertion into the IP header. A TCP that serves multiple users must be able to identify which to test and update, or for the Smart Market, which to insert.

An implementation needs to address, among other things, how to facilitate interprocess communication between the ECI and the communication software, and allow a study of perturbations to the currently fine-tuned transport protocol (especially TCP) operation.