Value-centric Information and Intelligence Sharing Scheme

2.1 Main objective

Here, we define the main objective of VCN. First, we define the network components as follows. We denote a piece of information by I_i, where i is the identifier of the information in the network system. Its value is denoted by v_i. The simplest objective of VCN is to maximize the sum of v_i in the network system. We assume that v_i is composed of multidimensional attributes because information has various value aspects.

Furthermore, the network must use some resources, e.g., bandwidth and electricity, when transmitting and receiving information. Let r_i denote resources that I_i uses, and let R denote all of the resources available for use. Like v_i, r_i can consist of multidimensional attributes. We assume that there are n pieces of information in the network system and define the VCN objective by the following formula.

This is a well-known NP-hard problem called the knapsack problem. Although no algorithm is known to be the fastest for knapsack problems, Dynamic Programming (DP) can solve them quickly in practice. A greedy algorithm can give a suboptimal solution to this problem in the practical way.

2.2 Architecture

An overview of the architecture for the VCN concept is shown in Fig. 2. There are three domains: users, networks, and services. All layers have resource restrictions. In the user layer, a user’s device has resource restrictions such as the device’s battery power, central processing unit (CPU) performance, or network bandwidth. With these resources, users get information of value V_U. When users get information, networks must consume resources—mainly bandwidth and electricity. Networks convey information of value V_N, which is distributed without exceeding the resource restrictions. Services also consume resources to provide information. In doing so, services send information of value V_S and consume resources, such as bandwidth, electricity, and computing capacity.

Fig. 2. Overview of VCN architecture. Big ovals represent layers requiring value maximization. Small ovals represent maximizing interfaces.

VCN is intended to maximize all three values V_U, V_N, and V_S. However, there may be conflicts where different behaviors maximize different values. For example, when there are two users, A and B, V_U is represented by A’s max value (ΣV_UA) plus B’s max value (ΣV_UB). However, if the network resources are insufficient to satisfy the requests of both users, then the information requested by one of the user’s, either A or B, is not forwarded; this behavior maximizes V_N but not V_U.

2.3 Maximization methods

In this subsection, we describe how to get the optimal combination of information transferred. As previously mentioned, we assume that v is one of the information metrics (e.g., the popularity of the information) and is a cardinal number and that r is one of the resources, e.g., bandwidth. One of the simplest approaches is to use a greedy algorithm. The greedy algorithm for knapsack problems selects the information combinations sorted by v_i/r_i. We show one of the simplest examples in Fig. 3. This example shows the user getting some information, which is shown in the information table, from a service via the network. The network has restricted resources represented by R_N = 20. Examples of this restriction include bandwidth. First, we use the information set {I₁, I₂, I₃, I₄, I₅}. With this information set, the combination that maximizes the utilization of resources is {I₁, I₂, I₃}. On the other hand, the combination that maximizes the network value is {I₄, I₅}. These results are obviously different.

Fig. 3. One of the simplest examples of VCN. A user is receiving several information streams from a service, and network resources are restricted to R_N =20.

The greedy algorithm sorts the information set by v_i/r_i and returns its answer {I₄, I₅} as a solution. The procedure for obtaining this solution is similar to quality-of-service (QoS) classification: QoS uses priority, and we can use v/r priority here for the greedy algorithm. However, it is not always true that the greedy algorithm can get the optimal solution.

If we add information I₆ to the above information set, the greedy algorithm selects only {I₆} because its v_i/r_i is higher than those of others; this is not an optimal solution. As previously mentioned, DP is well known as one method for solving knapsack problems. DP selects {I₄, I₅} as the optimal solution. However, its calculation cost is high because it uses a matrix table having n rows and R columns. If the number of information bits is n, then the order of the greedy algorithm is 0(n), while the order of DP is , where g is resource granularity. For VCN, this cost might be excessive. R_N, which may be bandwidth, might be a big value, such as 100 Gbit/s or 1 Tbit/s. In addition, the amount of information ranges from several bytes to several gigabytes. This indicates that if we use DP to get the optimal solution, g must be small and the DP table must be extremely large.

3.1 Overview

In order to evaluate the VCN concept and conduct a practical experiment, we implemented a VCN proxy server (VPS) in a system. With this system, we assumed that all connections are controlled by VPS at the network entrance. We assumed that VPS has information about the network’s available resources. VPS controls the hypertext transfer protocol (HTTP) connections by assessing their values and resource consumption. Below, we describe VPS and the experimental results.

3.2 Implementation of VCN proxy server (VPS)

We implemented VPS as a Java program. VPS maximizes the value of information passing through it. We used a greedy algorithm to maximize information value because DP’s calculation cost is too high. VPS behaves as follows.

When a user sends an HTTP request to VPS, VPS accepts the connection and creates a proxy connection instance. The proxy connection instance includes two connections: Client Side and Server Side. The user’s HTTP request is read by Client Side. If it contains a host entry, Server Side tries to connect to the server identified in the host entry. If the connection is successful, Server Side writes the user’s HTTP request to the server and reads the server’s HTTP response. However, that read operation is restricted by the bandwidth controller. If the read size exceeds the restriction set by the bandwidth controller, Server Side suspends the read operation. Under the restriction, Server Side forwards the HTTP response to Client Side, and Client Side sends the HTTP response to the user. Server Side and Client Side have trash queues for storing packets forwarded from the server. A connection manager analyzes packets in the trash queue and controls the bandwidth according to the connection’s value (v) and the resource (r) using each connection’s bandwidth controller. Value v is defined in a value table, which stores tuples of a regular expression for the universal resource locator (URL) and value v. Each connection is evaluated against the value table by using its content’s URL. Resource r is defined by its content length, and the connection manager sorts the connections by v/r. The top P_v% of all connections at a given moment are allocated R_v% of the bandwidth. P_v and R_v can be set independently to any value. For example, setting P_v = 50 and R_v = 50 yields normal proxy server usage.

5.1 Value definition

In this article, we used the term information value many times. However, we avoided the fundamental question of “What is the information metric?” Defining the information metric is a difficult and controversial problem. We think that there is no universal solution that will suit everyone. Our assumption is that the definition will be grounded in two fundamental principles: objectivity and cardinal number. From this standpoint, some examples of information metrics are given below.

Objectivity: The information metric should be based on engineering factors and thus not be subject to the feelings or opinions of individuals. We assume that there are two aspects: objectivity and subjectivity. Using measures associated with subjectivity prevents comparisons because they are deeply related to individual feelings. For example, subjective metrics include the beauty of movies and photos, the appeal of music, and human interest in information. By contrast, the resolution or bitrate of movies, photos, or music and the popularity of content based on the number of requests for it are objective metrics.

Cardinal number: The information value should be a quantity, i.e., a numerical value. When characterizing information by a number, we need to consider two types of number: cardinal numbers and ordinal numbers. A cardinal number expresses a numerical quantity (count) while an ordinal number indicates a ranking. An example of a cardinal value is information popularity. Popularity can be expressed by the number of people who like it, and this is a cardinal number.

5.2 Feasibility

In achieving the VCN concept, there are two main difficulties. One is the technical problems posed by VCN’s need for deep packet inspection and by solutions to the knapsack problem. The calculation costs might be high enough to cause network services to suffer delay. Actually, the network performance of VPS is only 24 Mbit/s, despite its fully asynchronous operations. We believe that higher-grade computers and more sophisticated programs will run VCN at wire speed; nevertheless, electricity consumption will still be problematic.

The remaining difficult issue is “Who decides the information metric and how?” Because numerical information values are essential in VCN, the metric should be decided carefully. For example, if only the network provider determines the metric, the result may not be acceptable to users and service providers. For this issue, we think that it is important to have transparency in the valuation process. For example, users and services might request the network to provide their own values, and networks would run the value-maximizing process on the basis of those values.