The Value of Intermediaries in Network Commerce by Wai Kiong Chong Andrew Jennings E-mail: [email protected], [email protected] Computer Systems Engineering Royal Melbourne Institute of Technology GPO Box 2476V Melbourne, Victoria 3001 Australia

Current version: February 28, 1995 Pre-publication in The Fourth International Conference for Young Computer Scientists (ICYCS'95) July 19-21,1995 Beijing P.R. China

Keywords. ATM Networks, Markov chain, Brokers, Agents, Advertising, Network cost.

1. Introduction There is at present a small volume of commerce on networks such as the Internet. We expect that this will grow rapidly over the next year with the availability of mechanisms for the exchange of credit card information. Most of the exchanges will take place via the World Wide Web and will typically involve consumer items rather than large purchases such as business equipment. In this paper we examine the future possibilities for network based commerce with a particular emphasis on the establishment of open markets. Is it possible to create highly efficient international markets based on network commerce? This study examines the place of intermediaries in network commerce and opens up some questions concerning how to foster intermediaries, or at the other extreme how to regulate the number of intermediaries. A key question is how well the existing Web scales to a much larger market. If we project Internet growth rates then we can expect the number of users to double every six months. Our modelling and simulations show that intermediaries can significantly improve efficiency and that we should look seriously at encouraging a growth in the number of intermediaries. The growth in intermediaries can in part alleviate the difficulties of scaling network commerce by several orders of magnitude over the next year or two. We have to be careful when we talk of efficiency in the context of network transactions. Here we take an economic view: in any process of purchasing there is a need for coordination and exchange of information. A buyer may examine a catalog and similarly a seller may circulate an advertisement. As long as the process of coordination is cheap in comparison to the price of the product or service then we can call this an efficient marketplace. Here we are concerned with exactly how efficient we can make a network based marketplace. So if a transaction consists of the passing of an advertisement, followed by a query and finally by delivery of a product (eg. a file or video) across the network then we are concerned only with the advertising and query. We are only concerned with the issues of coordination and overheads. If network overheads are much smaller than competing methods of sale then there may be a migration of commerce onto the network.

1

As a prototype for delivery of information and services, the Web is very powerful in encouraging the use of network services. With most users of the network being charged on an institutional basis rather than by individual billing there are few mechanisms to limit traffic. Cocchi et al [Cocchi 1993] discusses charging regimes for Internet use, and MacKie-Mason and Varian [MacKie 1994] give an overview of the economics of the Internet. In this paper we do not discuss the impact of different charging regimes: we are only concerned with means of coordination and organisation for improved efficiency. We consider an ATM based network, and in principle our results are relevant to any network services based on this backbone. It is interesting to contrast the design principles of Usenet news with the World Wide Web. Usenet news was constructed at a time when bandwidth was a very scarce commodity, and as a consequence it has quite sophisticated distribution mechanisms that minimise the use of bandwidth. Distribution of the news follows a hierarchy of delivery nodes in such a way that users effectively share the costs of distribution. The Web is at the other end of the spectrum: it uses bandwidth entirely on an individual basis. It makes very little attempt to share the costs of distribution amongst users. The use of intermediaries is one important means of re-organising the Web to reduce coordination costs. Most importantly no fundamental changes to the Web protocols are required, only a charging regime that encourages the growth of intermediaries. In principle the Internet does not carry broadcast advertising. Instead suppliers promote their services through other media (eg. newspapers) and as a result interested users look at Web pages. The impact on efficiency is however identical to the case of broadcast distribution of advertisements. If five users each retrieve an advertisement then the same network bandwidth is consumed as if the advertiser had sent the message to those users. In this paper we will primarily refer to the broadcast mode of advertising as this is more familiar, but it is important to note that our results on efficiency are relevant to the current Internet mode of operation. We consider two types of intermediaries: brokers and agents. A broker acts on behalf of sellers in searching for suitable purchasers. It is an intermediary for collecting advertisements and circulating to users. In a similar fashion an agent acts on behalf of purchasers in searching for suitable products. Although in theory it is possible to combine the functions in a single point, it

2

is preferable to separate the functions so that each party maintains a degree of trust in their selected intermediary. The structure of the paper is as follows. In Section 2 we introduce the formulation of the network model that determines the effective capacity of links based on an assumed offered traffic. We also present our assumptions for routing and an example network for evaluation. In Section 3 we address the question of how sellers and buyers meet in the network, and evaluate the case where there are no intermediaries. In Section 4 we introduce network intermediaries and evaluate their effectiveness. Section 5 discusses the results of our simulations, and finally we present our conclusions and suggestions for further work. Some of the issues we have not considered include the cost of machine processing time in servicing requests, and the internal costs of distribution of information. We have narrowed our focus to the consideration of the cost of sending messages by international networks. At some time in the distant future these costs may be a small part of the overheads, but at present this link and message costs are a very significant part of the cost. For local networks within a building or a city then quite different considerations may be important.

2. Model Formulation In this section, we present an ATM network model formulation that consists of many sellers and many buyers. All sellers are distributing advertisements to buyers. The network model can estimate the required bandwidth for the advertisement distributions in each network link. Throughout the paper, we assume that the total network cost is directly proportional to the required bandwidth. We also assume charging is independent of distance in the network. To begin with, we briefly describe the topology and the routing algorithm in the model. For the traffic source, we are using a computational approach of estimating the equivalent bandwidth of each source.

3

Topology In the beginning, we construct a network model that could represent the Asia Pacific region (15 nodes) and define a network that consists of 50 nodes at major centres. (Figure 2.1) Every node in the network can be either a single buyer or a single seller. It can be also a regional location that consists of multiple sellers or buyers. If we introduce any broker or agent, he can connect to any node. In other words, the links between the nodes can represent the backbone links of the B-ISDN network. We will not calculate the link capacity within a local centre (e.g. Singapore). In practice, there is a certain administration cost for sending information within a local centre.

TOKY

BEI J SEO

HK NEWD

BKK

TAI

VI ET MAL SI N

JKTA

DAR

SYD MELB

AUCK

Figure 2.1 An ATM network in Asia Pacific Region. Routing There are various routing techniques implemented in current networks such as TYMNET, TRANSPAC ARPANET, SNA and DNA. These networks use various types of shortest path algorithms that route packets from source to destination over a least-cost path. They mainly differ in cost criteria used (Onvural 1994). We will not discuss the cost criteria in this paper. For routing purposes, an ATM network can consider as a packet-switched connection-oriented network. In this paper, we assume an ATM network and the routing algorithm used is Floyd's shortest path algorithm (Floyd 1962). Although all network designers can design a good routing algorithm for the network, they can not ensure that each seller will find his nearest buyer to

4

advertise to or each buyer will find his closest seller (shop) to buy their products. (In real life, you might buy a teddy bear in a foreign country but it is on sale at a shop near your home.)

Traffic Source Each traffic source in the real world is variable. There is an extensive research literature devoted to estimating the required bandwidth of a stochastic traffic source in the network. Guerin et al. proposed a simple approximation for the equivalent capacity or bandwidth requirement of a single or multiplexed connection on the basis of their statistical characteristics (Guerin 1991). His computation is the combination of two different approaches; one based on a fluid-flow model and the other on an approximation of the stationary bit rate distribution. Elwalid and Mitra offer the computation of the effective bandwidth of a general Markovian traffic source (Elwalid 1993). In this paper, we use this general Markovian source to estimate the bandwidth requirement for each traffic source. In order to estimate the required bandwidth for each seller, we propose an example of a seller distributing multimedia advertisements to potential buyers. The advertisement consists of 5 seconds of video with CD-audio quality of sound, 10 seconds of voice annotated high resolution image and 20 seconds of voice annotated text. (Table 2.1) This is considered representative of advertisements on a commerce oriented network.

State Type of information 1

video/audio

2

voice annotated high-

3

Mean bit rate in Mbps Mean Holding Time in s 1.484

5

resolution image

*0.512

10

voice annotated text

0.0323

20

*2000x2000 resolution, 12 bits/pixel

Table 2.1

An example of a multimedia advertisement

5

In a long run average, we assume that each seller is a generic Markovian source with multiple states correspond to different bit rates. Figure 2.2 is the state diagram of a Markov model describing the above example.

1.484Mbps:5s

video/audio State 1

0.5

0.5 0.5

0.5 0.5

voice/image

voice/text State 3

State 2

0.5 0.0323Mbps:20s

0.512Mbps:10s

Figure 2.2 Markovian state diagram for advertisement

Generic Markovian Traffic Source λ ) where M is the Each traffic source is considered as a fluid source characterised by (M,λ infinitesimal generator of a controlling Markov chain. The source generates fluid at the constant rate λs when in state s. For a given buffer size B and the overflow probability p, it can derived the equivalent bandwidth e, of a Markovian source is approximately equal to the maximal real eigenvalue of an essentially non negative matrix [Λ Λ - (1/ζ)M] where Λ = diag(λ λ ) and ζ = log(p)/B. λ ) = g(λ λ) e(λ where g(λ) denotes the maximal real eigenvalue of [Λ Λ - (1/ζ)M] . In order to form the infinitesimal generator matrix M of the markov chain, we need to know the transition probability for each state.

Transition Probability If pij is the probability that the source is transferring from state i to state j, we can write the pij in matrix form:

6

p11 ..... p1m  P = (pij) =

:

:



pm1..... pmm 

We assume that the probabilities of transition from each state to the next state are evenly distributed. It means P12, P13,P21,......P32 are all equal and is 0.25 for the state diagram shown in Figure 2.2. We also assume that each node does not have an empty queue at any time (during the specific interval of studies). Since each international node is a concentration of local traffic, we assume that there is a steady stream of advertisements of the type outlined above.

P =

0

0.5

0.5

0.5

0

0.5

0.5

0.5

0 

Long term behaviour of the transition probability When we raised the matrix P to some power n, the matrix Pn has all its elements strictly positive. We called such transition probability matrix P as regular. Formally, for a regular transition probability matrix P, we have the convergence: lim Pijn = πj > 0 for j = 0, 1, . . . . . N, n->∞

If we have n = 10, the matrix Pn converges to:

0.333

0.333

0.333

P10 = 0.333

0.333

0.333

0.333

0.333

0.333

7

In the long run, 33.3% of the time is spent each state. We can also approximately calculate the mean source rate by considering the long run behaviour of the Markov model. Since we assume that the mean holding time for each state is 5, 10, and 20s respectively, the relative proportion of time spent in each state will be 5(0.333), 10(0.333) and 20(0.333) accordingly. So the mean source rate can be estimated by : λ = (1.484*5*0.333 + 0.512*10*0.333 + 0.0323*20*0.333) (5*0.333+10*0.333+20*0.333) = 0.3767 Mbps.

Bandwidth and Network Parameters Elwalid also observed that the effective bandwidth decreases monotonically with the increase in ζ from peak source rate λ when ζ = -∞ to the mean source rate λ when ζ = 0. Figure 2.3 shows the equivalent bandwidth of the above example decreases from the source peak rate 1.484 Mbps to the mean source rate 0.378 Mbps. In our model, we assume that the buffer size B = 48 Mbits (6 Mbytes), and the overflow probability p = 1x10-5, which gives us the equivalent bandwidth to be about 0.8 Mbps.

B andwidth and Network P arameters 1.6

Peak R ate = 1.484

1.4 1.2 1 0.8 0.6

Mean R ate = 0.378

0.4 0.2 0 -100

-10

-1

-0.1

-0.01

0

log(p)/B

Figure 2.3 The equivalent bandwidth varies with the network parameters

8

3. How do seller and buyer connect? To begin our analysis, we construct a model that consists of only sellers and buyers. This model is to simulate an environment without any broker and agent. Even though it is a simple model, there are many scenarios for the buyers and sellers to connect to each other. The model is based on the assumption that buyers and sellers have the freedom to choose whom to connect to. For example, each seller can randomly connect to any buyers disseminating an advertisement as in Figure 3.1. Buyer Buyer Seller Buyer Seller Buyer Seller Buyer Buyer

Figure 3.1 Random connection between buyers and sellers

In an open market, the network should have no restriction for buyers to receive advertisements from any seller who wish to sell their product or services. Some might choose the nearest available node to connect. Others might choose the furthest node to connect especially if the network is "free". We setup the following three scenarios to study sensitivity: 1) each buyer will receive an advertisement from the nearest seller; 2) each buyer will receive it from the furthest seller; 3) each buyer will receive it from an average distance seller. Table 3.1 shows that if all buyers received advertisement from the furthest node, the network cost will increase by about 45 to 50%* from the network cost of the average distance connection case. If all buyers choose to connect to the nearest node, then the cost will reduce to * the network cost that we calculated is not based on distance charging.

9

about 60 to 70%. For a long run behaviour analysis, we construct a network model by assuming each buyer will randomly pick a seller or choose not to connect at all. We observed that the random cases have a mean corresponding to the average distance connection. (Refer to Figure 3.2) These results show a very significant cost impact if each seller and each buyer always need to connect to a further node to find something. In a free market situation, we can not control sellers and buyers to be cost conscious if the network cost is "free". The situation will be worse if we have millions of buyers and sellers entering into the future "electronic purchase world" at the same time.

No of Node

Closest

Furthest

Average

Random Avg

15

15.44

63.32

44.78

37

50

85

419

281

274

Table 3.1 Comparison of network cost between various connection scenarios

450 400

AAAA AAAA AAAA AAAA AAAA AAAA AAAAAAAA AAAAAAAA AAAA AAAA AAAA AAAAAAAA AAAAAAAAAAAA AAAAAAAAAAAA AAAAAAAAAAAA

350

Random

300 AAAAAAAA AAAA AAAA AAAA AAAA AAAAAAAA AAAAAAAA AAAA AAAA AAAA AAAAAAAA AAAAAAAAAAAA AAAAAAAAAAAA AAAAAAAAAAAA 250

Random Average AAAA AAAAAAAA AAAAAAAA AAAAF urthes t

200 Nearest

150

AAAAAAAAAAAA Average

100 50 0 1

3

5

7

9

11

13

15

17

No of random runs

19

21

23

25

No of S eller = 10 No of Buyer = 40

Figure 3.2 Network Cost for Various Connection Scenarios

10

Full Connection At one extreme every seller could connect to every buyer and supply an advertisement (Refer to Figure 3.3). In the real world, it is rare to have full connection, especially if there are millions of buyers and sellers on the network. However, this is a good scenario to make comparison in the simulation environment because it is an environment where the maximum amount of information is being distributed to the buyers. We run a model that consists of 10 sellers advertising their new arrivals to all 40 buyers at the same time. The total cost is 2,687 units. Buyer Buyer Seller Buyer Seller Buyer Seller Buyer Buyer

Figure 3.3 Full connection between buyers and sellers

4. Intermediaries Before we go into how the intermediaries could connect to the sellers and buyers, let us first define the role of a broker and an agent. In this paper, the role of the intermediaries is different from the definition by Schwartz in his Resource Discovery model (Schwartz 1993). A broker is to act on behalf of sellers and advertise services or information that are of use to others on the network. He can be the information provider for sellers to reach out to an agent that acts on behalf of the buyers or to buyers directly. A broker can also be a desktop publisher that helps a client to advertise or to act on behalf of the seller to provide services or information to their buyers. In the future, a broker can also collect essential information from agents to learn about buyers' needs. So in the

11

long run, a broker can then provide more specific advertisement or information to a specific individual agent or buyer. The role of an agent is to act on behalf of buyers and to maintain a list of services or information that circulated by brokers in the form of catalogue or list of advertisements. Once he is contacted by the buyer, he will start to seek out information requested by the buyer. In the case of full connection or broadcast model, the agent will automatically send out all information received from the brokers to the buyers that are connected. In the future, an agent could also be intelligent enough to perform user modelling (Jennings 1992). When there are a few thousand information providers, it makes sense to limit the number of advertisements sent to buyers since they may put up "No junk mail" signs if there is excessive advertising. There are many ways of introducing the brokers and agents between the seller and the buyers. Figure 4.1 shows a proposed model where each broker and agent only serves their clients within a region. We assume that all sellers search for their nearest broker for advertising service, and all agents will only serve the buyers within the region. In order to have all sellers' advertisements sent to each individual buyer (equal amount of information to the full connection model), we need to have all brokers disseminate the information to every regional agent.

12

Buyer

Buyer

Agent Seller

Buyer

A

Seller

Broker Buyer

Seller Seller

Buyer

Agent

Seller

B

Seller

Broker

Seller Seller

Buyer

Seller

Buyer

Seller

Buyer

Agent

Buyer

Figure 4.1 All brokers contact every agent

It can be argued that it is not necessary that sellers will always look for nearest broker and buyers will contact the closest agent for their services. In the real world, it is also not possible to restrict sellers connecting to any brokers or buyers to any other agents. Sellers can choose to send the advertisements to a regional broker or some dedicated brokers. Buyers can also choose to receive advertisement from any available agent, or their dedicated agents. If we assume full exchange of information between brokers and agents then there is no additional advantage for a buyer connecting to another agent compared to their regional agent since each agent is carrying complete information. For the proposed model, we assume each seller and buyer chooses their nearest available broker and agent for service. In the future, if the price of choosing a nearer broker or a closer agent is much cheaper than a further broker or agent, it is reasonable to assume that each seller and buyer will be more cost conscious in selecting a broker or an agent. To construct a model with intermediaries, we use the topology of the full connection model describe in the last section. We increased the number of brokers and agents to study the sensitivity of the network cost. Although the model is capable of simulating advertisements each with a different mix of multimedia information, in this paper, we assume each seller is

13

using about 0.8 Mbps bandwidth for their advertisement distribution. Since we already assume the network cost is directly proportional to the bandwidth usage, we can assume each seller requires 0.8 unit of network cost. Refer back to Figure 4.1 for the typical situation with 2 brokers. If there are 4 sellers send their advertisement to broker A and 6 sellers to broker B, then broker A and broker B will require 3.2 units and 4.8 units of the network cost to distribute to each agent respectively. Since we need all advertisements distributed to every buyer, each agent will require 8.0 units of network cost to disseminate to all buyers. We have also assumed that when a new advertisement becomes available it is communicated to brokers, and in turn it may be then communicated to agents. Coordination between brokers and agents is then needed to maintain a complete list of current advertisements that are available for local distribution. If all of the information changes frequently then there is little advantage in storage. However it is difficult to anticipate exactly what proportion of the advertisements will change over time. We have conservatively estimated that the update traffic volume is only 20% of the total.

How many brokers and agents? Figure 4.2 shows a sharp decrease of the network cost when we introduce an agent. There is a 36% reduction in the network cost when we increase the number of agents from 2 to 6. When we increase the number of agents further from 6 to 10, the reduction of the network cost is below 10%. In our model we did not include the fixed asset or capital cost for each agent and broker. Nevertheless it is important to take note of the existence of such capital cost here. If we assume that the overheads and other fixed cost for introducing agents are about 10% or more, then it is not worthwhile to increase the number of agents further. In this case, we can conclude that 6 agents is the optimum number. Figure 4.3 also shows there is no significant advantage if we increase the number of brokers from 2 onwards. It is reasonable with only 10 sellers in this model. The cost incurred between agent and buyer can be borne by either agent or buyer themselves. The cost incurred between broker and agent is strictly administration cost which none of the buyers or sellers will like to bear. It is good to keep it as low as possible.

14

3000 " full connection" without broker & agent = 2687 units

2000

S eller-Broker B roker-Agent Agent-Buyer T otal

1000

0 0

2

4

6

8

10

12

14

16

18

No. of Broker = 2

No. of Agent

Figure 4.2 Network Cost Distribution with 2 Brokers

15

Network Cos t f or D iff erent No. of Agent 4000 No of Broker 3000 1 2000

5 10

1000

0 1

3

5

7

9

11

13

15

17

No. of Agent

Network Cos t with D if ferent No of B roker 4000 No of Agent

3000

2 2000

6 12

1000

0 1

3

5

7

9

11

13

15

17

No. of B roker

Figure 4.3 Network cost with different number of Agents and Brokers.

16

5. Discussion We have presented results for a simulated network that roughly corresponds to a regional network of ATM links supporting a high volume of network commerce. In order to simulate the situation we have to adopt some simplifying assumptions, so we should be careful in interpretation. In our network model we have assumed that all advertisements are similar in type and correspond to the multimedia mix outlined in Section 2. In determining the link capacity there is an inherent assumption of steady-state behaviour on the network. We cannot determine peak loads or transient behaviour with this type of model. With these qualifications there are some clear trends that emerge from the simulations. The introduction of intermediaries results in dramatic savings in link costs, and can play a major role in encouraging network commerce. Our results indicate up to 70% savings in network costs with the use of intermediaries based on local connections. In cases where local brokers and agents are heavily used in preference to distant sellers the savings are substantial. We should again emphasise that we have not used a distance based charging model in our simulations: in this model link bandwidth from Sydney to Melbourne is charged at the same rate as the equivalent bandwidth between Sydney and Hong Kong. If distance based charging is included then we would expect even more dramatic savings. The Internet Resource Discovery Project [Schwartz 1993] has emphasised the use of brokers and this formed the inspiration for this study. However there are subtle differences in the definition of brokers and agents as considered in the Resource Discovery Project and our current study. Our focus is on the process of coordination of buyers and sellers which has some important differences from the resource discovery model. The exchange of advertisements we have assumed is generally equivalent to a multimedia message or a Web page with this content. In our model the brokers and agents actually store and forward these advertisements rather than just establishing pointers to their network location. Another key difference is that we have assumed an ATM backbone network, which we believe will give our results more lasting value than simply considering the current Internet technology.

17

We have assumed that rather than being utility functions that are maintained by the network operator brokers and agents are instead viable services in their own right. Brokers act on behalf of sellers and agents act on behalf of sellers, and we would expect that perhaps they will maintain their operations by taking a percentage of each transaction as they do in other areas of commerce. Our results show that brokers and agents can effect sufficient savings to become self-sustaining, although this is of course dependent on the charging regime. In one sense the Internet does not support broadcast advertising so our simulations could be considered hypothetical. However the current growth of the Web shows that there is a form of advertising growing very rapidly. Users who place Web pages tend to advertise their presence through other media such as Usenet news and newspapers. Perhaps one day soon we will see the first Web page advertisement on a television commercial. If a large number of users then access the Web page then this is equivalent in the use of network bandwidth to the broadcast mode of advertising, at least in principle. For the purposes of simplicity in the paper we have assumed normal advertising rules apply and have considered a broadcast mode of operation. There are two network technology improvements that may influence our conclusions. Widespread availability of multicast distribution could make broadcast distribution of advertisements much more efficient, and would reduce the network cost for almost all of our simulations. However the use of multicast can also be of similar assistance to broker/agent communication so we would contend that there will still be advantages in the adoption of intermediaries. We are yet to simulate this situation so we cannot suggest that the ratios would be similar. Similarly, widespread adoption of caching could significantly alter our results. In a cached system if a user retrieves a Web page then for a certain time this page will be held at a local server so that other users can access the local version directly, saving the international link bandwidth for transferring the page again. For commonly used pages the caching mechanism can produce quite substantial improvements in efficiency. There is an immediate difficulty in charging here since perhaps the first user to retrieve the page may bear the cost and later users benefit. Of course there is the possibility for brokers and agents to act as the caching site. This has the possible advantage that the brokers and agents could financially support the storage

18

necessary for caching, rather than the network operator bearing this cost. Since we have not simulated caching in our model we cannot make a definite judgement on this combination. There is potentially a deeper problem in the adoption of intermediaries. Experience on the Internet [Schwartz 1993] for the case of ftp traffic shows that there is a very strong tendency to ignore local mirrors and instead to retrieve information from the distant site. This behaviour does not seem to have altered with the availability of tools such as archie that identify sites with the information. It is apparent that this is a challenge for sites that wish to establish themselves as brokers or agents. Since one of our goals is to maintain an open market where free exchange of goods is possible, we do not want to restrict users to use the local agent. In some cases the best price or the best service may come from a very remote site. This can especially be true for highly specialised products or services. We can only speculate that with the correct charging regime combined with creative intermediaries that it may be possible to alter this seemingly intractable behaviour. Certainly the savings possible with the use of intermediaries should be available to further this cause.

6. Conclusions and Further Work We have simulated an international network of the near future with an emerging market of network commerce. The use of agents and brokers as advocated by other workers in resource discovery shows very substantial savings in network costs. It is clear that if we wish to encourage an efficient market we should foster the growth of intermediaries. In the present charging regime where users are not charged for actual bandwidth usage but are instead charged on an unlimited usage basis it will be difficult to encourage the use of intermediaries by users. However it is still possible to give discounts to information brokers and agents. If we wish to encourage efficiency then we should at least offer network usage at a discount rate to brokers and agents. In future charging regimes it may be possible also to encourage users to adopt the use of local agents and brokers where possible. We plan to study the interaction between charging regimes and network cost in a similar fashion to the study by Cocchi et al [Cocchi 1993 ]

19

References Cocchi, R., Shenker, S., Estrin, D., Zhang, L., 1993 'Pricing in Computer Networks: Motivation, Formulation, and Example' IEEE/ACM Transactios on Networking, vol.1, no.6, pp 614-627. Elwalid, A.I., Mitra, D., 1993 'Effective Bandwidth of General Markovian Traffic Sources and Admission Control of High Speed Networks', IEEE/ACM Transactions on networking. vol.1, no.3, pp.329-343. Floyd R.W. 1962 'Algorithm 97' (Shortest Path), Communication of Association for Computing Machinery, vol.5, pp.345. Guerin, R., Ahmahi, H., Naghshineh, M., 1991 'Equivalent Capacity and Its Application to Bandwidth Allocation in High-Speed Networks', IEEE Journal on selected areas in Communications, vol.9, no.7, pp.968-981. Jennings, A., Flower, M., ., 1992 'A Multimedia Shop' Proceedings of the International Interactive Multimedia Symposium Jan, pp.573-583 MacKie-Mason, J.K., Varian, H.R., 1994 'Some Economics of the Internet' Tenth Michigan Utility Confernece at Western Michigan University. March 25-27, 1993. Onvural, R.O., 1994, Asynchronous Transfer Mode Networks: Performance Issues, Artech House , Boston, London.Nicholson,P Schwartz, M., 1993 "Internet Resource Discovery at the University of Colorado" IEEE Computer,Sept vol 26, no.9. pp 25-35

20