26
Page | 1 Incorporating Congestion Control in BGP considering its Economic and Policy Effects Dhaval Malnika Jay Kothari Kishan Parekh Vinita Shah TLEN 5700: Research Methods April 25, 2014 Advisors: Prof. Jose Santos Prof. Mark Dehus Special Thanks: Jay Robertson Senior Manager, IP/CDN Software Engineering Internap

Team 13 - Jay Kothari - Apr 25 2014 457 Pm - Group 13 Final Capstone Paper

  • Upload
    an-uj

  • View
    14

  • Download
    2

Embed Size (px)

Citation preview

Page | 1

Incorporating Congestion Control in BGP considering its

Economic and Policy Effects

Dhaval Malnika

Jay Kothari

Kishan Parekh

Vinita Shah

TLEN 5700: Research Methods

April 25, 2014

Advisors:

Prof. Jose Santos

Prof. Mark Dehus

Special Thanks:

Jay Robertson

Senior Manager, IP/CDN Software Engineering

Internap

Page | 2

ABSTRACT

The Internet can be visualized as an interconnection of Autonomous Systems (ASes), which use

the Border Gateway Protocol (BGP) as the standard for exchanging network information among

each other. The growth of the Internet has made it possible to access information with ease, but

on the other hand, it has given rise to the problem of congestion. BGP does not have an inherent

mechanism to avoid or detect congested areas and routes packets through paths that are often

congested. The AS_PATH attribute in BGP is an indicator for selecting the best path to reach a

certain Internet destination. Current congestion control practice involves manual network

configuration changes by network engineers. Our proposal is to amend the BGP protocol to

incorporate congestion control mechanism. Thus, we have developed a tool that continuously

monitors a link for congestion and automatically carries out AS path prepending to re-route the

traffic from a congested link onto another, thereby improving network reachability and reducing

packet loss. A new model discussing the related economic and policy consequences is also

discussed in the paper.

Page | 3

Table of Contents

Serial No. Introduction Page No.

I.

i.

ii.

Introduction

Statement of the problem

Research Question

5

5

6

II. Literature Review 9

III. Research Methodology 10

IV. Research Results 19

V. Discussion of Results 22

VI. Conclusion and Future Research 24

VII. References 25

List of Figures

Serial No. Name of Figures Page No.

Fig. 1.2.1 ISP 2 incurs cost for ISP 1’s traffic 8

Fig. 3.1 Congested BGP topology 11

Fig. 3.2 Congestion Detection algorithm 13

Fig. 3.3 Congestion Recovery algorithm 14

Fig. 3.4 Congestion-controlled BGP 15

Fig. 3.5 Co-ordination between Tier 1 ISPs 16

Fig. 3.6 Multihomed customer 17

Fig 3.7 Economic incentives for Tier 2 ISPs to form a full-mesh 18

Page | 4

Serial No. Name of Figures Page No.

Fig 4.1 Bandwidth vs. % Loss due to congestion 20

Fig 4.2 Comparison of Bandwidth vs. % Loss 20

Fig. 4.3 Graph of transmit load on interfaces 21

Fig. 4.4 Flow analysis graph 22

Fig 4.5 % Loss with a congested alternate path 22

Page | 5

I. Introduction

i. Statement of the problem

The Internet is comprised of a large number of ASes and exchanging routing information

between two or more ASes is achieved using BGP. BGP is a vector distributed routing protocol

and uses TCP as its underlying mechanism, making it reliable [1]. It is not only important for the

BGP sessions to be reliable, but also scalable to provide inter-domain Internet connectivity.

However, in case of link congestion, the BGP router still forwards packets along the same path,

increasing packet loss and degrading network reachability. It not only leads to the packets being

dropped and re-sent repeatedly, but also affects cost, performance, and adds delay which

eventually affects the customer.

Current solutions to cope with congestion majorly focus more on increasing the available

bandwidth by adding more links between routers [2]. It is not economical for ISPs to meet traffic

requirements by incrementing bandwidth at Layer-1 as traffic is irregular and not always at

maximum capacity. Service Providers rely on mechanisms such as Multi-Protocol Label

Switching – Traffic Engineering (MPLS-TE) and route manipulations that can control

congestion to a certain extent by redirecting traffic through suboptimal paths within an AS, but

for inter-AS congestion-control, there is no clear technique. Thus, a multi-provider congestion

control mechanism needs to be implemented by enhancing BGP, to reduce overall packet loss

and improve reachability.

Considering the previous reasons, there is a need to incorporate an intelligent BGP

implementation that will choose a path considering the network utilization and traffic patterns at

some point in time. We propose the use of AS Path Prepending as a method to reduce congestion

in BGP. By prepending AS_PATH to the destination prefix, we deflect all traffic flows for that

Page | 6

destination onto another suboptimal link, which helps solving delay, packet loss, and throughput

problems.

Along with BGP, there come certain types of relationships between different ASes. They

are commonly referred to as peer, transit, and customer relationships. These three types have

traffic agreements between them, what is more commonly known as Service Level

Assurances/Agreements (SLAs). These SLAs in customer-transit relationships have monetary

arrangements. Generally, peer-peer relationships are settlement-free or require certain financial

agreements depending upon the shared number of routes between both the ASes. For example,

Century Link and Verizon have settlement-free peering, since both have equal number of routes

to share, while Verizon and Cogent have paid-peering since the number of routes shared from

both ISPs have a significant difference. Tier-I ISPs have more number of Internet routes as

compared to Tier-2 ISPs. Tier-I ISPs offer transit services to Tier-2 ISPs and Tier-2 ISPs offer

transit services to its customers.

Alternate links to a destination might not have the same financial agreements. Deflecting

traffic onto these alternate links, which are not congested, could be a concern to the customer-

transit and peer-peer relationships. This traffic deflection could cause problems in the SLAs.

ii. Research Question:

How will integrating a congestion-control mechanism in BGP make it a network efficient

protocol and what will be the economic and policy implications as well as feasibility from an

industry standpoint?

The rest of the paper is organized as follows: The following sub-section provides the sub-

problems associated with BGP congestion. Section II highlights the literature review and the

research done so far regarding BGP congestion. Section III discusses the research methodology.

Page | 7

Section IV provides the results we achieved after running tests on a congested network topology.

In Section V, we discuss on the results achieved. Section VI proposes the future scope for this

research and conclusion.

Research Sub-problems:

This section will discuss critical issues and challenges to the possible adoption of

congestion control mechanism in BGP.

• When diverting traffic onto an alternate link, if the alternate path to a destination

becomes congested as well, there is a possibility of path flapping or ‘bouncing effect’.

The traffic will flap between the best path and the alternate path until all the data is sent

to the destination. This behavior has to be avoided to prevent packets arriving out of

order or packets not arriving at all at the destination. Packet losses will cause

retransmission requests through the flapping paths, which will degrade network

performance further. Bouncing effect can be eliminated by the router having the best path

to the destination by refusing to divert traffic again if it receives the traffic back from the

alternate link.

• An ISP implementing congestion control mechanism may divert traffic to some other

ISP. However, some of the ISPs may restrict such traffic flow and as a result the traffic

will be black holed. For example, if ISP 2 has implemented congestion control

mechanism and if it experiences congestion, then the BGP packet sent by ISP 1 towards

ISP 2 will now have a different path towards the destination. Assuming the traffic is now

flowing through ISP 3, and if ISP 3 has implemented policies for the traffic to be

blocked, all the traffic being diverted to ISP 3 will be dropped. As a result, we might have

to implement this mechanism globally and SLAs between the ISPs will have to be

Page | 8

modified. Traffic could be offloaded between the ISPs on an equivalent basis. This also

highlights the importance for the ISPs to collaborate among one another for such a

congestion control mechanism.

The inclusion of these modifications into BGP can lead to economic and policy challenges.

• Economic: As shown in Fig 1.2.1, congestion occurs on the link between ISP 1 and

Transit Provider. In order for ISP 1 to reach its destination, sending data over ISP 2’s link

may not seem a plausible solution for ISP 2 because it may not want traffic of a customer

on its own link from which it does not obtain profits. This can be a problem when the

available alternate paths of ISP 2 are via the transit provider. ISP 2 might have to pay the

transit provider, which can result into an overall increase in their costs, which will

eventually be borne by the customers or certain financial arrangements will have to be

made between the ISPs. Also, such methods may not be acceptable to ISP 2, which are

having their bills increased due to diversion of traffic.

Fig. 1.2.1 ISP 2 incurs cost for ISP 1’s traffic

• Policy: By default, ISPs give highest priority to traffic from their customers, followed by

peers, and then transit. The alternate or suboptimal path might be via either peer or transit

and if congestion control in BGP will be implemented, the peering and traffic

arrangements rules will change to a certain extent. Currently, certain ISPs block traffic of

Page | 9

other ISPs from entering their domain unless they are explicitly permitted. This

restriction needs to be changed in order to successfully implement congestion control

mechanism in BGP. In addition, ISPs may be required to share internal topological

information and/or their available routes with other ISPs, which can lead to ISPs losing

their competitive edge [3]. As a result, ISPs might become hesitant to adopt this idea into

BGP.

II. Literature Review

When BGP was invented in 1989, there was no support for traffic congestion based on

network load, cost and performance which led to numerous instances of Internet outages. Also,

during congestion, BGP still prefers the “congested best-path route” to a given destination.

Experts have done extensive research to overcome this frailty in BGP. The key research ideas

related to our research are listed below.

Gupta et al., MacKie-Mason and Varian proposed a new Internet model with users being

the customers, and providers as suppliers of services like videos, stock quotes, etc. Their model

deals with priority pricing mechanisms, where the users decide how much priority they want to

give to their own traffic, before sending it towards the destination. In this way, when congestion

occurs, lower priority traffic can be dropped and the users with higher priority traffic can pay a

higher amount than usual [4]. This research gave us the idea of enabling BGP to prioritize traffic

based on its distance (maximum AS path length) from the congested area.

Fujinoki, Wang and Gao proposed ideas to include multiple paths in the BGP routing

table. They devised methods like Multi-Path BGP, D-BGP, and B-BGP, which added suboptimal

paths into BGP for a given destination without creating routing loops [5] [6]. This increased overall

network efficiency and throughput in case of congestion.

Page | 10

Sebakor et al. proposed a unique method of introducing a control system for each AS,

which controls the inter-domain traffic. This method reserves capacity on a link for certain IP

traffic and allocates bandwidth to other traffic automatically. The results obtained from this

research showed that the control system could achieve efficiency in link utilization, avoid

possible link congestion, and decrease the packet loss rate [7]. The concept of a ‘BGP Path

Announcer’ in our research stems from this idea.

According to statistics from the AT&T backbone, about 30% of routes are AS prepended

which indicates the significant impact of the attribute on the entire Internet domain [8].

Prepending an AS path does not compromise the BGP routing resilience or increase the BGP

routing table size. Rocky K.C. Change et.al proposed various advantages of AS path prepending

approach for means of traffic engineering [9]. We incorporated such an approach into our

algorithm to reduce manual network configuration and enable predictable control over change in

traffic flow.

Agarwal et al. proposed an Overlay Policy Control Architecture (OPCA), which runs as

an overlay network on top of BGP in the Internet. OPCA allows an AS to make route changes to

other remote ASs and achieve quicker fail-over. It also provides an AS the capability to control

the inbound traffic, and thereby help in avoiding congestion [10]. This idea cannot be

implemented in real world because traffic of one AS is being manipulated by another AS, which

would violate current norms in the Internet. To overcome this limitation, we created a model

where each AS is independent but requires collaboration with other ASes.

III. Research Methodology

We adopted the following methodology for our research:

Step 1: Proving that BGP cannot handle congestion

Page | 11

We simulated a non-congestion-controlled BGP topology in a multi-AS environment.

The topology consisting of five ASes had all the possible paths from the source to the destination

and selected the best path using the BGP decision process. Thereafter, we simulated a congested

BGP scenario by continuously flooding stream of data using Iperf, thereby congesting a

particular link between two ASes.

Fig. 3.1 Congested BGP topology

Iperf is a free Linux based network performance tool used to measure various TCP and

UDP parameters such as bandwidth, delay, jitter, packet loss [11]. Here, client is the sender and

server is the receiver. All ASes in Fig. 3.1 have external BGP (eBGP) running between them. We

created data streams of different bandwidths ranging from 1 Mbps to 100 Mbps and streamed the

data from Ubuntu client to Ubuntu server. Also, we observed the % packet loss and % transmit

load on the outgoing interface of AS 2 for different bandwidths.

Step 2: Logging and analyzing the data

We logged the data from the results and analyzed various performance parameters like

average end-to-end delay, average packet loss rate, and average load. Then we processed the

Page | 12

logged data to plot the graph of Bandwidth vs. % Loss and Bandwidth vs. % Load to compare

them with the results of congestion-controlled BGP scenario.

Step 3: Developing the algorithm

We developed a tool to achieve congestion control and route optimization using AS Path

Prepending approach. The tool is basically divided into two components.

1. The Initial Setup: Simple Network Management Protocol (SNMP) is extensively used

to poll the router for various information regarding the interface index, IP addresses, the

next hop addresses, routing table attributes and information about the BGP routes, and

the AS path lengths. A link object is created that maps the link index to its directly

connected neighbors and possible routable network destinations. A dictionary mapping of

IP addresses to their AS path lengths is also stored during the initial setup.

2. Active Continuous Measurement: The tool actively monitors all the links of the router

for its utilization values. The formula used for calculation link utilization is: [12]

Input Utilization = ∆𝑖𝑓𝐼𝑛𝑂𝑐𝑡𝑒𝑡𝑠 × 8 × 100

(𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑠𝑒𝑐𝑜𝑛𝑑𝑠 𝑖𝑛 ∆) × 𝑖𝑓𝑆𝑝𝑒𝑒𝑑

Output Utilization = ∆𝑖𝑓𝑂𝑢𝑡𝑂𝑐𝑡𝑒𝑡𝑠 × 8 × 100

(𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑠𝑒𝑐𝑜𝑛𝑑𝑠 𝑖𝑛 ∆) × 𝑖𝑓𝑆𝑝𝑒𝑒𝑑

where ∆𝑖𝑓𝐼𝑛𝑂𝑐𝑡𝑒𝑡𝑠 and ∆𝑖𝑓𝑂𝑢𝑡𝑂𝑐𝑡𝑒𝑡𝑠 represents the count of inbound and outbound

octets of traffic respectively.

A variable threshold value is defined to identify congestion on the link. When the

link utilization value exceeds the threshold value, the tool starts its congestion control

mechanism.

Page | 13

Fig 3.2 Congestion Detection algorithm

• From the link object, the interface index, IP addresses along with their AS path lengths,

and it’s directly connected interfaces are retrieved. In case the link has never faced

congestion, a subset list of IP addresses having AS path length equal to the maximum AS

path length is obtained.

Page | 14

• Route-maps to prepend the AS path are entered into the router, which triggers in its

immediate neighbors a BGP decision process to choose an alternate best path for the

concerned IP addresses. Route maps can also be entered on the basis of Netflow statistics

to prepend the largest traffic flow through the router. Such a decision is left onto the ISP.

• The program then returns to monitoring link utilization. In case the link utilization value

still persists above the threshold value, the router again starts the congestion control

mechanism. Since, it has already prepended the AS value for the maximum possible AS

path lengths, it now prepends the AS path for IP addresses with the next maximum AS

path length.

Fig 3.3 Congestion Recovery algorithm

Page | 15

If the link utilization value decreases beyond a minimum threshold value, the link is no

longer congested. At this point, the ISP will try to restore its previous best forwarding path. The

tool identifies from the link object, the minimum of the AS path value for which route maps were

entered along with the corresponding IP addresses. The tool then continues to monitor the link

utilization.

Fig 3.2 also takes care of the bouncing effect or “flapping” of routes. After diversion of

traffic, in case no peers have the required capacity to allow the traffic, the traffic is routed back

to ISP, which initially prepended the path. At this stage the ISP can drop/allow the traffic.

Step 4: Implementing the algorithm

We implemented the mechanism explained in Fig. 3.2 and Fig. 3.3 in the following BGP

scenario, where the link between AS 1 and Switch 0 is 10 Mbps and the link between AS 2 and

AS 3 is 100 Mbps. Thereafter, we analyzed the traffic patterns to compare average end-to-end

delay, average packet loss rate, and average load between traditional BGP and congestion-

controlled BGP.

Fig. 3.4 Congestion-controlled BGP

Page | 16

Step 5: Interviews

As per our topology and obtained results; we conducted interviews with experts from

ISPs to know their viewpoints over a possible integration of congestion-control mechanism into

BGP. The interviews helped us analyze the industry standpoint on the possible integration of this

new concept in BGP. The interviews mainly focused on these major domains:

• Awareness: How far the industry experts were aware about the BGP congestion problem

and any existing mechanism that could solve the issue

• Feasibility: The feasibility of incorporating the algorithm in regards to technical and

commercial viability

• Prospective clientele: Based on the current market scenario, who among Tier 1 ISPs, Tier

2 ISPs, and customers, would have an interest in implementing such automatic

congestion control mechanism

Step 6: Proposed Model

Considering the current business relationships among the ASes, we propose a modified

Internet model to address the economic and policy sub-problems mentioned previously. The

model is divided into the following scenarios:

Fig. 3.5 Co-ordination between Tier 1 ISPs

Page | 17

Scenario 1: Explicit Co-ordination among the Tier 1 ISPs

As addressed in the economic and policy sub-problem, diversion of traffic onto another

link will not be acceptable to an ISP due to increase in costs and trust issues. However, we

suggest that all Tier-1 ISPs engage in coordinated inter-domain routing and traffic engineering as

shown in Fig. 3.5. As of today, ISPs take independent decisions to carry out its own network

engineering, which can impact another ISP's traffic. Coordinated routing on the other hand,

requires exchange of policies and routing decisions allowing predictable incoming and outgoing

traffic engineering control. Rahul et al, further explain how ISPs can end in a win-win situation

while maintaining limited information disclosure among the ISPs. We support the idea of ISPs

engaging themselves in a two-way exchange of information and route negotiation between the

upstream and the downstream routers as it provides a strong base for global deployment of our

tool. With the ability to predict traffic changes, ISPs can divert traffic without worrying about

local resource policies and system instability, thus optimizing cost and network performance [13].

Fig. 3.6 Multihomed customer

Scenario 2: Mutihomed ASes and Customers

"Mutihomed" means connected to more than one Internet service provider as shown in

Fig. 3.6. Over the past few years, there has been an increase in the growth of multihomed ASes

Page | 18

and customers to improve network connectivity and performance. We propose the use of our tool

for all such enterprise networks that have more than one connection to the Internet. The tool will

allow automated traffic engineering control by varying the AS's traffic among the different

service providers it is connected to, thus providing better network utilization. Such automated

load balancing abilities allow the AS to support more users connecting to the Internet leading to

increased revenues.

Fig. 3.7 Economic incentives for Tier 2 ISPs to form a full-mesh

Scenario 3: Economic arrangements between full-meshed Tier 2 ISPs

According to the current Internet model, the interconnection between Tier 2 ISPs is sparse as

compared to the interconnection between the Tier 1 ISPs [14]. Tier 2 ISPs purchase transit from

the Tier 1 ISPs in order to access the Internet. Congestion on the Tier 2 transit link results in

more costs for the Tier 2 ISP in order to allow the additional traffic. We suggest in our model

that all the Tier 2 ISPs in a specific region have full mesh connectivity between themselves. Such

complete peering benefits the Tier 2 ISPs by reducing the overall transit cost. Considering the

transit cost between Tier 1 and Tier 2 is $1/Mbps, the Tier 2 ISP facing congestion can offload

Page | 19

the additional traffic through another Tier 2 peer for the duration of congestion. The costs of the

additional traffic can be billed:

• At the same cost as it would have been to a Tier 1 ISP or,

• Based upon specific peering and traffic load balancing agreements among the Tier 2 ISPs

• Based upon pay-per-use of network resources.

Such model allows support for inflow and outflow of money among the Tier 2 ISPs as against

the current unidirectional outflow of money from a Tier 2 ISP to a Tier 1 ISP. Another additional

benefit is that the Tier 2 ISP has a virtual link capacity equal to the entire Tier 2 full mesh

bandwidth, thus allowing support for more customers and traffic generating income for the Tier

2 ISP.

IV. Research Results

After implementing the BGP topology as shown in Fig 3.1, the first step was to prove that

BGP cannot handle congestion. We performed the following steps:

1. Injected a stream of data packets on a serial link for the bandwidths ranging from 1 Mbps

to 100 Mbps from the Ubuntu client to the Ubuntu server

2. There was no packet loss from 1 Mbps to 1.544 Mbps. For various link bandwidths from

1.544 Mbps to 100 Mbps, we observed that the packet loss increased exponentially and

remained constant (97%) after 70 Mbps as shown in Fig. 4.1.

Page | 20

Fig. 4.1 Bandwidth vs. % Loss due to congestion

3. Fig. 4.2 shows the comparison graph of Bandwidth (BW) vs. % Loss before and after

implementing the congestion-control algorithm. The blue curve represents a flow of

traffic within the link capacity of Fa0/0 (AS1). Thus, when the BW was between 1 and 10

Mbps the corresponding loss was 0%. As we increased the bandwidth beyond 10 Mbps,

the % loss increased exponentially, as shown by the red curve. At 60 Mbps, we started

our congestion control mechanism and as seen, the traffic gets diverted onto the Fa0/1,

AS2 link and the flow does not suffer from loss of packets and reaches the destination.

Fig. 4.2 Comparison of Bandwidth vs. % Loss

Page | 21

4. Fig. 4.3 shows the comparison graph of % transmit load on outgoing interfaces Fa0/0 and

Fa0/1 for routers AS1 and AS2 in Fig. 3.4 respectively. The values are plotted before

congestion as well as before and after implementing the congestion-control algorithm.

The bar graph shows that the transmit load was 1% before congestion. After congestion

on the Fa0/0 interface of AS1, the transmit load rises to 63% i.e. the interface is using

around 63% of the available transmit bandwidth. The red portion in the bar graph

represents % packet loss on the Fa0/0 interface. Once the algorithm diverts the traffic via

an alternate path, the transmit load gradually decreases to 19% on Fa0/0 of AS1 and

increases to 30% on the interface Fa0/1 of AS2.

Fig. 4.3 Graph of transmit load on interfaces

5. Fig. 4.4 compares the % Loss for 3 flows, which have the same best path to reach the

destination, ultimately leading to congestion of the link. Referring Fig. 3.4, flow 1 is

traffic from customer connected to AS 1, while flow 2 and flow 3 represent external

traffic passing through AS 1. Due to congestion there is a 70-80% loss for flow 1 and

flow 2. Since the link is completely congested, there is 100% packet loss for flow 3. We

Page | 22

implement the algorithm for the external traffic, but not for our customer traffic. When

the algorithm is implemented, the packet loss for all the three flows becomes 0%.

Fig. 4.4 Flow analysis graph

6. The Iperf result shown in Fig. 4.5 is to illustrate the bouncing effect. It shows that link

Fa0/0 (AS1) has 68% packet loss when the link is congested. The packet loss on the link

reduces to 0% because of the AS Path Prepending. However, the alternate link also gets

congested diverting the traffic back to the original path. Thus, we see an increase in

packet loss from 0%, back to 68% on the Fa0/0 link.

itplab@itplab-Inspiron-530:~$ iperf -s -p 6005 -i 1 –u

[ 3] local 10.2.0.2 port 6005 connected with 10.0.0.2 port 48495 [ ID] Interval Transfer Bandwidth Jitter Lost/Total Datagrams [ 3] 72.0-73.0 sec 1.14 MBytes 9.57 Mbits/sec 0.093 ms 1736/ 2550 (68%) [ 3] 73.0-74.0 sec 1.14 MBytes 9.57 Mbits/sec 0.102 ms 1737/ 2551 (68%) [ 3] 74.0-75.0 sec 3.43 MBytes 28.8 Mbits/sec 0.049 ms 624/ 3073 (20%)

Congestion Management

[ 3] 75.0-76.0 sec 3.58 MBytes 30.0 Mbits/sec 0.044 ms 0/ 2551 (0%) [ 3] 76.0-77.0 sec 3.58 MBytes 30.0 Mbits/sec 0.038 ms 0/ 2551 (0%)

Redirection via another path

[ 3] 155.0-156.0 sec 1.14 MBytes 9.57 Mbits/sec 0.088 ms 1737/ 2551 (68%) [ 3] 156.0-157.0 sec 1.14 MBytes 9.57 Mbits/sec 0.108 ms 1737/2551 (68%)

Peer had no available bandwidth

Fig. 4.5 % Loss with a congested alternate path

V. Discussion of Results

Page | 23

1. As per Fig. 4.1, the rate at which the data is being sent is much higher as compared to the

actual capacity of the link resulting in packet loss. The BGP protocol continues to send

traffic over the congested link even though there are alternate paths to reach the

destination. This shows that BGP has no inherent congestion control mechanism.

2. From Fig. 4.2, we see that link bandwidth was insufficient to handle the incoming traffic

resulting in packet loss. As the link utilization increased above the threshold value, the

algorithm diverted the additional traffic via an alternate link, which had enough capacity

to forward the incoming traffic towards the destination at 0% loss.

3. Fig. 4.3 shows that the link congestion results in packet loss and a high transmit load on

the interface. Since the algorithm diverts traffic via an alternate link, the loads get shared

among the interfaces.

4. The Iperf result in the fig. 4.5 shows that our algorithm takes care of situation when the

alternate path to the destination is also congested. Diverting the traffic causes congestion

on the alternate link which then sends the traffic back through the original path. Even

though the traffic bounces back with no reduction in packet loss, the traffic is not

completely dropped or black holed. However, if such a scenario occurs the ISPs can

decide whether to allow/block the traffic.

5. To understand the industry standpoint and the feasibility of the tool, we interviewed

industry experts from Tier 1 and Tier 2 ISPs. We asked them about their awareness of the

BGP congestion problem, deployment feasibility and viability, and potential customers

who could buy our solution. The experts from CenturyLink suggested that the

effectiveness of the tool depends on the acceptance of the Open Internet model by other

ISPs. They suggested making the tool capable of recognizing multiple links between two

Page | 24

Tier 1 ISPs and pass traffic through those links. They also agreed that such automatic

congestion control tool can be readily deployed for multihomed ASes and customers.

Tier 2 ISPs supported the idea as this solution not only solves their congestion problem,

but also benefits them and their customers economically. They also have an added

advantage in the form of inflow and outflow on money, and an increased customer base.

VI. Conclusion and Future Research

In this paper, we have proposed a novel approach for controlling congestion by diverting

traffic onto another link by means of AS Path Prepending traffic engineering approach. Initially,

our tool creates a database of all the necessary information, actively probes the links to detect

utilization, and then undertakes necessary steps to control congestion. The entire process is

automated as opposed to the current ad-hoc engineering methods and does not have any

additional dedicated hardware or software requirements. The paper also proposes a model that

supports an Open Internet system and explains various scenarios for which the tool can be

readily deployed. From the results obtained, we conclude that such automated tools are the future

for network operations management and troubleshooting.

Our current BGP congestion control mechanism tool works on three simple principles of

link utilization monitoring, BGP AS Path Prepending attribute, and co-ordination among the

ISPs. The future scope of the tool is based on the decision of the ISPs and their willingness to

adopt the Open Internet model. The Open Internet architecture provides the ability to monitor

links of different ASes and divert traffic based on least link utilization. The ISPs can choose to

develop additional complex features to increase the tool’s efficiency to measure delay and

throughput characteristics and also take into consideration Netflow statistics before diverting

onto an alternate path.

Page | 25

VII. References

[1] G. Siganos and M. Faloutsos, “Analyzing BGP policies: Methodology and tool,” in

INFOCOM 2004. Twenty-third AnnualJoint Conference of the IEEE Computer and

Communications Societies, 2004, vol. 3, pp. 1640–1651.

[2] I. van Beijnum, “BGP: Chapter 6: Traffic Engineering,” Building Reliable Networks with the

Border Gateway Protocol, Sep-2002. [Online]. Available:

http://oreilly.com/catalog/bgp/chapter/ch06.html. [Accessed: 06-Dec-2013].

[3] N. Hu, P. Zhu, H. Cao, and K. Chen, “Routing Policy Conflict Detection without Violating

ISP’s Privacy,” in Computational Science and Engineering, 2009. CSE’09. International

Conference on, 2009, vol. 3, pp. 337–342.

[4] A. Gupta, D. O. Stahl, and A. B. Whinston, “A stochastic equilibrium model of internet

pricing,” May-1997. [Online]. Available:

http://www.sciencedirect.com/science/article/pii/S0165188996000036. [Accessed: 23-Apr-

2014].

[5] H. Fujinoki, “Multi-path BGP (MBGP): A solution for improving network bandwidth

utilization and defense against link failures in inter-domain routing,” in Networks, 2008.

ICON 2008. 16th IEEE International Conference on, 2008, pp. 1–6.

[6] F. Wang and L. Gao, “Path diversity aware interdomain routing,” in INFOCOM 2009, IEEE,

2009, pp. 307–315.

[7] M. Sebakor, N. Theera-Umpon, and S. Auephanwiriyakul, “Centralized control system in

interdomain routing environments,” in International Symposium on Intelligent Signal

Processing and Communications Systems, 2008. ISPACS 2008, 2009, pp. 1–4.

Page | 26

[8] H. Wang, R. Chang, D.-M. Chiu, and J. C. Lui, “Characterizing the performance and

stability issues of the AS path prepending method: taxonomy, measurement study and

analysis,” in Proceedings of ACM SIGCOMM Asia Workshop, 2005.

[9] R. K. C. Chang and M. Lo, “Inbound traffic engineering for multihomed ASs using AS path

prepending,” IEEE Netw., vol. 19, no. 2, pp. 18–25, Mar. 2005.

[10] S. Agarwal, C.-N. Chuah, and R. H. Katz, “OPCA: robust interdomain policy routing and

traffic control,” in 2003 IEEE Conference on Open Architectures and Network

Programming, 2003, pp. 55–64.

[11] “Iperf - The TCP/UDP Bandwidth Measurement Tool.” [Online]. Available:

http://iperf.fr/. [Accessed: 23-Apr-2014].

[12] “How To Calculate Bandwidth Utilization Using SNMP - Cisco,” 26-Oct-2005. [Online].

Available: http://www.cisco.com/c/en/us/support/docs/ip/simple-network-management-

protocol-snmp/8141-calculate-bandwidth-snmp.html. [Accessed: 23-Apr-2014].

[13] R. Mahajan, D. Wetherall, and T. Anderson, “Towards coordinated interdomain traffic

engineering,” in Proceedings of Third Workshop on Hot Topics in Networks (HotNets-III),

2004.

[14] “The Global Internet Peering Ecosystem.” [Online]. Available:

http://drpeering.net/AskDrPeering/blog/articles/The_Internet_Peering_Ecosystem__The_Tie

r_2_ISP.html. [Accessed: 23-Apr-2014].