Hisham Iqbal Roll # MS Final Project Report Topics in Internet Research (CS 678)

Transcription

1 Project Title: IGP - Traffic Engineering Hisham Iqbal Roll # MS Final Project Report Topics in Internet Research (CS 678) Project Final-Report Topics in Internet Research (CS 678) Page 1 of 1

2 Abstract To support the requirements of multimedia and other state-of-the-art applications and provide the end user with required services Traffic Engineering techniques need to be applied to the currently deployed infrastructure to improve it efficiency and performance. The paper talks about the objectives of traffic engineering its requirements and related work. It tries to explain motivation behind Traffic Engineering and reviews some of the currently used Traffic Engineering methodologies (and it predecessors) within an Autonomous System. (Intra-domain Traffic Engineering). Some IETF projects related to Traffic Engineering are also touched upon. Later it discusses the classification of the Traffic Engineering System. In the last part of the paper Traffic Engineering with an Autonomous System is discussed i.e. intra-domain Traffic Engineering or Interior Gateway Protocol Traffic (IGP) Engineering. IGP Traffic Engineering is based upon the optimizing the underlying network without deploying any new routing methodology such as MPLS. Two main routing protocols in this aspect (IGP Traffic Engineering) are Open Shortest Path First (OSPF) and Intermediate System to Intermediate System (IS-IS). General Optimal Routing is the best deployment of the traffic on the underlying network which can be easily achieved using MPLS. It has been concluded at the end that IGP traffic engineering can perform within a few percent of the Optimum General Routing. A number of approaches to support this argument are discussed later in the paper. Project Final-Report Topics in Internet Research (CS 678) Page 2 of 2

3 Table of Contents 1 INTRODUCTION TRAFFIC ENGINEERING CONGESTION CONTROL TRAFFIC ENGINEERING STEPS TO FOLLOW TRAFFIC ENGINEERING PERQUISITES INTRA-DOMAIN VS. INTER-DOMAIN ROUTING INTRA-DOMAIN TRAFFIC ENGINEERING ROUTING PROTOCOLS TRAFFIC ENGINEERING OVERVIEW OF METHODOLOGIES ADAPTIVE ROUTING / DYNAMIC ROUTING TYPE OF SERVICE ROUTING EQUAL COST MULTI-PATH ECMP NIMROD OVERLAY MODEL CONSTRAINED-BASED ROUTING INTEGRATED SERVICES DIFFERENTIATED SERVICES MPLS CLASSIFICATION OF TRAFFIC ENGINEERING SYSTEM TRAFFIC ENGINEERING IGP INTRA-DOMAIN ROUTING PROTOCOLS MAJOR APPROACHES TO TRAFFIC ENGINEERING Optimizing Existing Network The Overlay Approach Optimal Routing / Optimal General Routing (OPT) TRAFFIC ENGINEERING APPROACHES USING THE EXITING NETWORK EQUAL COST MULTI-PATH (ECMP) OPTIMIZED MULTI-PATH (OMP) Opaque LSA STATIC OPTIMIZATION USING LOCAL SEARCH HEURISTIC The Proposed Approach Project Final-Report Topics in Internet Research (CS 678) Page 3 of 3

4 5.4 WEIGHT MANAGEMENT BY SOME EXTERNAL ENTITY The Centralized Approach Approaches to Assigning Link Weights DYNAMIC OPTIMIZATION USING ADAPTIVE RANDOM SEARCH Packet Drop Probability The Adaptive Random Search Algorithm INTEGRATED APPROACH TO TRAFFIC ENGINEERING The Objective function ENHANCEMENT TO STATIC LOCAL SEARCH HEURISTIC APPROACH CONCLUSION REFERENCES Project Final-Report Topics in Internet Research (CS 678) Page 4 of 4

5 1 Introduction Over the last decade the Internet has developed into an essential communications media, supporting all types of activities ranging form simple and web browsing to video conferencing and video on demand etc. The delivery of Internet communications services has therefore become a very challenging task. End-users and applications are demanding very high QoS support; therefore performance optimizations of large scale IP networks, especially the Internet backbones, have become an important problem. To meet the demanding requirements of the end user and high-tech applications the Internet must convey traffic from source to destination in an efficient, reliable and economical manner. The focus of this report is on the currently prevailing Intra-domain traffic engineering techniques i.e. traffic engineering techniques used within a given Autonomous System (AS) in the Internet (IP based networks). 1.1 Traffic Engineering Traffic Engineering is referred to as improving user performance and making more efficient use of network resources; adapting the routing of traffic to the prevailing demands and routing the traffic in a manner to best utilize the available resources without causing any interference to other users [1]. One of the most distinctive functions performed by Internet traffic engineering is the control and optimization of the routing function, to maneuver traffic through the network in the most effective way. 1.2 Congestion Control One of the major objectives in optimization of the IP based network is to minimize congestion. Network resource is considered congested if the arrival rate of the packets exceeds the output capacity of the resource over a certain period of time [1]. Following are some of the proposed methodologies for congestion control or avoidance:- Project Final-Report Topics in Internet Research (CS 678) Page 5 of 5

6 Congestion can be controlled using Long / Medium / Short time scales. Further details can be seen in [1]. Reactive (Preventive) congestion management policies react (take proactive action) on existing congestion problems to improve it. Supply side congestion management policies increase the effective capacity available to traffic in order to control congestion. Whereas Demand side congestion management policies regulate the offered traffic to lessen congestion. 1.3 Traffic Engineering Steps to Follow Following are some of the steps that a traffic engineering system must perform to optimize the performance of the underlying network. These steps need to be performed iteratively. First phase is to define the relevant control policies that govern the operation of the network. Second phase is a feedback mechanism involving the acquisition of some data from the operational network. Third phase is to analyze the network state. Fourth phase is the performance optimization of the network based upon the results obtained from the previous steps 1.4 Traffic Engineering Perquisites To engineer any network traffic one needs to know the following information about the network:- The amount of packets entering your network and The amount of packets leaving your network, It is preferable that some information about their destination is also known. For further detail please see [2]. Project Final-Report Topics in Internet Research (CS 678) Page 6 of 6

7 1.5 Intra-domain vs. Inter-domain Routing The main objective of Intra-domain routing (IGP) [2] is to select the best path available towards each destination based on some metrics (Delay, Bandwidth, Congestion etc) used inside an AS. The fundamental requirement here is that the IGP should be able to react quickly to changes in topology. The main objective of Inter-domain routing (EGP) [2] is to select the best path towards each destination that is compatible with the routing policies of the transit AS(s). The fundamental requirement here is that EGP should be scalable. 1.6 Intra-domain Traffic Engineering Most of the large scale IP networks use Interior Gateway Protocols (IGPs) such as OSPF (Open Shortest Path First), IS-IS (Intermediate System-Intermediate System) or RIP (Routing Information Protocol) that select paths based on some static criterion like link weight. These weights are typically assigned by the network operators. Routers use these protocols to exchange link weights and construct a complete table of the topology inside the AS [3]. As described in figure. Each router computes shortest paths and creates a table that controls the forwarding of each IP packet to the next hop in its path. Shortest path routing within an Autonomous System Project Final-Report Topics in Internet Research (CS 678) Page 7 of 7

8 1.7 Routing Protocols Routing protocols enable the routers to build routing tables based upon which the routing decisions are made from the source to the destination. Routing protocols are generally classified as either of the following.- Link State protocols, where each router explicitly advertises information to other routers in the network about the nodes to which it is connected. Routers then use a suitable path finding algorithm, such as Dijkstra's Shortest Path First algorithm [15], to work out where packets destined for a node should be sent. Implementation includes OSPF (OPEN SHORTEST PATH FIRST) and ISIS (INTERMEDIATE-SYSTEM INTERMEDIATE-SYSTEM) protocols for intra-as routing. Distance Vector protocols, where each router advertises the prefixes (or routers) which it currently knows how to reach along with an associated cost. Neighboring routers receive this information and update their own routing tables. Such protocols often implement a distributed shortest path first algorithm (Bellman Ford Algorithm) [14]. Implementation includes Routing Information Protocol (RIP). Link state protocols provide more functionality, and generally converge faster than distance vector protocols [4]. Consequently, link state protocols are usually considered preferable in situations where the increased memory requirements are not unaffordable [16]. 2 Traffic Engineering Overview of Methodologies In the following subsection prior work related to the Traffic Engineering of Internet or General IP networks (and its predecessors) is reviewed. 2.1 Adaptive Routing / Dynamic Routing In adaptive routing decisions were based on the current state of the network. It was used in APARNET. Project Final-Report Topics in Internet Research (CS 678) Page 8 of 8

9 Dynamic Routing uses distributed control to determine the paths that packets should take en-route to their destinations. The insufficiency of the legacy Internet interior gateway routing systems, mentioned above, is one of the factors that led to the developments of constraint-based routing methodologies like MPLS. 2.2 Type of Service Routing Here routing involves selection of different routes to the same destination with selection criteria as the Type of Service ToS field of an IP packet header. A more detailed description of this can be found in [5]. For Routing D T R 6 7 IPv4 Header Version Header Type of Service Total Length (Bytes) Length (TOS) ID 3-bit 13-bit Fragment Offset Flags Time To Live (TTL) Protocol Header Checksum Source IP Address Destination IP Address 31 D Low Delay T High Throughput R High Reliability Project Final-Report Topics in Internet Research (CS 678) Page 9 of 9

10 ToS-based routing is now out-of-date as the IP header field has been replaced by a Differentiated Services field. 2.3 Equal Cost Multi-Path ECMP This is one of the techniques that are used to address the deficiency in the Shortest Path First (SPF) interior gateway routing systems [6]. In the classical SPF algorithm, if two or more shortest paths exist to a given destination, the algorithm will choose one of them. The algorithm is modified slightly in ECMP so that the traffic between the nodes is distributed among the multiple equal-cost paths. One of the drawbacks of the ECMP is that load sharing cannot be achieved on multiple paths which have non-identical costs. Work related to the implementation of the global optimization of link weights and related issues with respect to ECMP are discussed in the later part of the paper. 2.4 Nimrod Nimrod is another routing system to provide heterogeneous service specific routing in the Internet, while taking under consideration multiple constraints [7]. This approach is not discussed in this paper because of it s very less or not deployment. 2.5 Overlay Model In this model, a virtual-circuit is created between the edge routers. Therefore two routers that are connected through a virtual circuit see a direct adjacency between themselves independent of the physical route taken. Thus, the overlay model essentially decouples the logical topology from the physical topology that the ATM, frame relay, or WDM networks use. Project Final-Report Topics in Internet Research (CS 678) Page 10 of 10

11 The Overlay Model Virtual Link Physical Link R2 R1 R2 s Routing Table R5 R4 R1: Link 1 R3: Link 2 R6: Link 3 R6 R3 The figure above shows a virtual circuit form router R2 to the other edge router R6, R3 and R1. Some of the issues with the overlay model are discussed in [8]. 2.6 Constrained-Based Routing It refers to the class of routing systems that compute routes through a network based upon certain constraints and requirements. Multi Protocol Label Switching (MPLS) is one such approach. In MPLS routing decision are based upon an additional tag that is assigned to each packet. This subsection reviews a number of IETF activities relevant to Internet traffic engineering. 2.7 Integrated Services This model requires resources, such as bandwidth and buffers, to be reserved a priori for a given traffic flow to ensure that the quality of service requested. Therefore the routers need to maintain certain state information for each flow. Project Final-Report Topics in Internet Research (CS 678) Page 11 of 11

12 Two services have been defined under the Integrated Services model: guaranteed service [10] (It is used for applications requiring bounded packet delivery time) and controlled-load service [11] (It is used for adaptive applications that can tolerate some delay but are sensitive to traffic overload conditions). The main issue with the Integrated Services model has been scalability [12], especially in large public IP networks since each router need to maintain state information of the flow. Integrated Services model requires explicit signaling of QoS requirements from end systems to routers [13]. The Resource Reservation Protocol (RSVP) performs this signaling function and is a critical component of the Integrated Services model. RSVP enables the sender, receiver, and routers of communication sessions to communicate with each other in order to set up the necessary router state to support the required service. 2.8 Differentiated Services The fundamental idea underlying the Differentiated Services (Diffserv) architecture is to categorize traffic into behavior aggregates, each behavior aggregate can then be treated independently of the other [1]. One of the reasons that led to the development of Diffserv effort is the scalability issue faced by the Intserv model. A Differentiated Services field in the IP header (DS field) has been defined in place of ToS field. Only the 0-5 bits are currently used and are called Diffserv Code Points (DSCP). Bits 6 and 7 are reserved for future use. User for Bits 0,1,2 Bits 3,4 Bit 6 (class) (drop precedence) EXP/ LU XXX XX 1 BE CSC 000 XXX AF1 AF2 AF3 AF ,10,11 01,10,11 01,10,11 01,10, Project Final-Report Topics in Internet Research (CS 678) Page 12 of 12

13 EF Two major classes shown in the table are Expedited Forwarding and Assured Forwarding. For an end-user of network services to receive Differentiated Services from its Internet Service Provider (ISP), it may be necessary for the user to have a Service Level Agreement (SLA) with the ISP. An SLA is a Traffic Conditioning Agreement (TCA) which defines certain rules or constraints for the traffic. SLA Architecture DS Domain Egress Ingress SLA SLA DS Domain SLA DS Boundary Node DS Interior Node SLA 2.9 MPLS MPLS extends the Internet routing model and enhances packet forwarding and path control [9]. Since MPLS is outside the scope of this project it is not discussed in the paper. Few other schemes like IP Performance Metrics, Flow Management and End Point Congestion Management are also proposed by different work groups. Details of which can be seen in [1]. Project Final-Report Topics in Internet Research (CS 678) Page 13 of 13

14 3 Classification of Traffic Engineering System Brief Classification of traffic engineering systems is given below further details of which can be seen in [1]. 1. In the time-dependent TE, historical information based on intermittent variations in traffic is used for routing plans and other TE control mechanisms. State-dependent TE uses the current state of the network to make routing plans for packets. 2. Since traffic engineering requires the computation of routing plans. The computation may be performed offline or online. Online computation is required when the routing plans must adapt to changing network conditions. Offline computation is done in cases where the routing plans need to be executed in real-time. 3. Centralized control has a central authority which determines routing plans and other TE control parameters on behalf of each router. In Distributed control each route determines route separately based on its own view of the current state of the network. 4. All traffic engineering algorithms require local or global network-state information. Local information refers to the state of a particular portion of the domain. Global information refers to the state of the entire domain undergoing traffic engineering. 5. Prescriptive traffic engineering evaluates alternatives and recommends a course of action. Whereas Descriptive traffic engineering, characterizes the state of the network and determines the impact of various policies without recommending any particular action. Project Final-Report Topics in Internet Research (CS 678) Page 14 of 14

15 6. Open-loop traffic engineering is the one in which the any control action does not use any feedback information from the current network state. Closed-loop traffic engineering mechanism is the one in which the control action utilizes feedback information from the current network state. 7. Tactical traffic engineering uses a tactical perspective to address specific performance problems. The Strategic traffic engineering approach uses an organized manner considering the long term consequences of specific action. 4 Traffic Engineering IGP In this section and in later sections Traffic Engineering only in terms of Intra-domain routing is considered (i.e. Traffic Engineering within an Autonomous System). Traffic Engineering (TE) as already defined as the task of mapping the flows of the traffic onto an existing underlying topology [19]. By mapping the traffic optimally over the network, problems caused by congestion and overloading of link can be avoided. Traffic Engineering Results in better performance of the same underlying network. IP routing typically uses shortest-path computation with some simple metrics such as hopcount etc to forward traffic form source to destination in a network. This methodology of IP routing based upon simple metrics does not make the best use of underlying network resources. 4.1 Intra-domain Routing protocols Open Shortest Path First (OSPF) is the most commonly used intra-domain internet routing protocol other commonly used protocol is Intermediate System to Intermediate System (IS- IS) protocol. In OSPF the traffic is sent to the destination using the shortest path(s). Shortest path from source to destination is selected based upon the link weights assigned to the links between the routers by some network operator or a system. The Project Final-Report Topics in Internet Research (CS 678) Page 15 of 15

16 assignment of the appropriate weights to the links between that routers may be based upon a number of factors (like link capacity, physical distance etc).in each router, the next link on all shortest paths to all possible destinations is stored in a table, a demand is sent to its destination by splitting the flow between the links that are on the shortest paths to the destination. In the traditional implementation of the OSPF routing protocol only a single path to the destination was used to send the traffic. 4.2 Major Approaches to Traffic Engineering There are two main approaches that are being used to take care of the traffic engineering problem in the internet (particularly within a domain) Optimizing Existing Network The basic idea is to optimize the existing network by carefully assigning link weights and then use OSPF routing protocol with these weights to route the traffic. Supporters of this type of methodology believe that traffic engineering requirements can be met without using and latest constraint based or overlay approach. In this paper most of the common methodologies / techniques used to implement Traffic Engineering using the existing network (existing network infra-structure and protocols) are considered The Overlay Approach With this approach logical connections are set up between the edge nodes (routers) and theses logical connections are overlaid to the existing physical topology. Routing decisions are made based upon these logical connections (virtual network) between the routers. These logical connections can be set up as ATM or Frame Relay [18]. Multi Protocol Label Switching (MPLS) is another method for setting up such logical connections. This approach is not constrained by the shortest path idea of routing. This approach is more favorable to the traffic engineering considerations. Some disadvantages Project Final-Report Topics in Internet Research (CS 678) Page 16 of 16

17 of using such an approach to traffic engineering are increased computation / communication overhead and potential instability of the network [19]. This approach though is widely implemented in the current Internet backbones has some scalability limitations. 1. To establish full meshed logical connections between edge nodes, each node has to set up logical connections to (N-1) other nodes. Therefore, N*(N-1) logical connections need to be established for the complete virtual network [18]. The complexity of the underlying network increases as the size of the network increases. More logical connections need to be established within a larger network. 2. Multiple logical connections usually go over the same physical link. Breakdown of a single physical link may cause multiple logical connections to fail, and may amplify the load. Traffic engineering requirements can be easily met if a constraint based approach like MPLS is used. This paper doesn t talk about these approaches to traffic engineering Optimal Routing / Optimal General Routing (OPT) It is the optimum optimization of the underlying network that can be achieved (using the overlay model). It is a benchmark against which different approaches to routing are compared. OPT can direct traffic along any paths in any proportions. It can establish one or more paths between every pair of nodes, and distribute arbitrary amounts of the traffic on each of the paths from source to the destination. 5 Traffic Engineering Approaches Using the Exiting Network Following are some of the proposed methodologies (techniques) for optimizing the existing underlying network (i.e. without using any overlay approach etc) to enhance performance:- Project Final-Report Topics in Internet Research (CS 678) Page 17 of 17

18 5.1 Equal Cost Multi-Path (ECMP) ECMP is an improvement upon the traditional OSPF/ IGP routing protocol to address the issues of traffic engineering. In the traditional Shortest Path First algorithm, if two or more shortest paths exist between the source and the destination, only one of them is chosen. In ECMP if two or more equal cost shortest paths exist between two nodes, the traffic between the nodes is distributed among the multiple equal-cost paths. The distribution of the traffic is performed in either of the following ways.- (1) Packet-based in a round-robin fashion, this approach can easily cause out-oforder packets. (2) Flow-based using hashing on source and destination IP addresses or any other IP header field. This approach depends upon the number and distribution of flows. Flow-based load sharing is generally effective in core public networks where the number of flows is large and heterogeneous. In ECMP, link costs are static and bandwidth constraints are not considered. ECMP distributes the traffic as equally as possible among the equal-cost paths without taking into account congestion status of each path. Another drawback of ECMP is that load sharing cannot be achieved on multiple paths with non-identical costs. Project Final-Report Topics in Internet Research (CS 678) Page 18 of 18

19 R1 Over Utilized Physical Link 50 % 50 % R5 R3 R2 R1 Unaware of R2-R4 link utilization R4 Equal Cost Multi-Path (ECMP) 5.2 Optimized Multi-Path (OMP) OMP is a compatible extension to OSPF. It utilities the Opaque LSA (IGP) to flood load statistics and proposes a way to adjust the forwarding of the traffic (i.e. moving traffic away form the congested paths). The traffic is distributed between equal cost multi-paths taking into consideration the congestion / utilization of the multiple links. With OMP traffic injection is varied across multiple paths based on link utilization as shown in the figure below. Project Final-Report Topics in Internet Research (CS 678) Page 19 of 19

20 R1 Over Utilized Physical Link 60 % 40 % R5 R1 aware of R2-R4 link utilization R3 R2 Opaque LSA Information Traffic distribution across R1-R2-R4 and R1-R3-R4 is not equal R4 Optimized Multi-Path (OMP) Opaque LSA An enhancement to the OSPF protocol has been introduced to support a new class of link-state advertisements (LSA) known as Opaque LSAs [20]. OMP use services provided by the opaque LSA to distribute traffic among multi-paths from source to destination. The information contained in Opaque LSAs may be used directly by OSPF or indirectly by some application to distribute information throughout the OSPF domain. Opaque LSAs consist of a standard LSA header followed by a 32-bit application-specific information field [20]. The Opaque LSA (like any other LSA) uses the link-state database distribution mechanism for flooding this information throughout the network. Opaque LSA has three different ranges of topological distribution within a network, referred to as Flooding Scope (link-local scope, area-local scope and throughout the Autonomous System). With OMP there is extra traffic overhead associated with the LSAs. Project Final-Report Topics in Internet Research (CS 678) Page 20 of 20

21 5.3 Static Optimization Using Local Search Heuristic In this approach [17] authors have addressed a very basic question. If any clever settings of OSPF weights can make it perform as well as General Optimal Routing and if yes then how well do OSPF routing perform on real networks? The answer to the first questions is that yes we can have a weight setting that would make OSPF perform close to the General Optimal Routing (authors have shown this for the proposed AT&T WorldNet backbone with their projected demands). We can have clever settings of OSPF weights that would make it perform within a few percent of the optimal general routing. The performance optimization if the OSPF weights depend mostly on the structure of the underlying network and the traffic demands of network. This contrasts the established view that OSPF routing leads to congestion and it is not flexible enough to take care of the traffic engineering (load balancing) demands of the underlying network. Authors, in this approach have shown that even for any randomly generated network topologies, 50%-110% more traffic demand can be met by using their optimal OSPF weight setting as compared to link weights setting based on some standard heuristic. In the proposed local search scheme if vector V denotes the vector of link weights V l that belong to the link set L. then to find a neighbor V` of V one of the following operation is performed [19]:- 1. Single weight change, where the link weight of only one link is changed to a value that has not been evaluated yet. 2. Evenly balancing flows, where equal cost paths are generated in order to balance the load. In a nutshell, weights of a subset of the links leaving a router R are changed to create equal cost paths for some destination t. A hash table mechanism is used to avoid revisits. Project Final-Report Topics in Internet Research (CS 678) Page 21 of 21

22 5.3.1 The Proposed Approach The general routing problem as defined in [17] is as follows. The underlying network is a multi-graph, G = (N, A). The nodes represent the routers and the arcs represent the links between the routers. Each arc a has capacity c (a). The capacity is the measure of the amount of traffic the link can take. We have a demand matrix D that shows how much traffic we need to send from source s to destination t. The demand, D(s, t) is zero if there is no path from s to t in G. Now for each non-zero demand D(s, t) we need to distribute the demanded flow over paths from s to t. There are no limitations to the distribution of the flows between the paths from s to t. The load l (a) on an arc a is the total flow over the link a. The main goal is to keep the loads l (a) within the capacities c (a). The cost function Φ sums the cost of the arcs. Φ = Σ Φ a (l (a)) Where for all a є A, Φ a (0) = 0 and Φ`a (x) = 1 if x є [0, 0.33c (a) [, Φ`a (x) = 3 if x є [0.33c (a), 0.66c (a) [, Φ`a (x) = 10 if x є [0.66c (a), c (a) [, Φ`a (x) = 100 if x є [c (a), [ The underlying idea behind the cost function Φ a is that it is cheap to send flow over an arc with small load. As the load l (a) approaches capacity c (a), it gets expensive and expensive to send flow over that link. When the load exceeds the capacity cost becomes huge. The objective function defined here is a piece-wise linear increasing and convex as shown below. Using this objective function the general routing problem can be solved optimally in polynomial time. Project Final-Report Topics in Internet Research (CS 678) Page 22 of 22

23 For link failure, one could pre-compute a suitable weight setting for each possible link failure for loading them when required. Time (0 10) Congestion (0 1) The Arc Cost Function Φ a (l (a)) In OSPF routing, for each arc a є A some weight is chosen. The weights chosen are thus used to determine which path(s) the traffic is to take from source to destination keeping in mind the considerations of the traffic engineering i.e. the cost function. Authors have used a local search heuristic to determine a weights vector (w a ) that minimizes the cost function Φ. Local search technique is an iterative procedure in which a neighborhood to the current node is selected keeping in mind that we have to minimize the cost function. The approach (descent method) is to select the best element from the neighborhood in each iteration and stop when a local minimum is found and there in no improvement in the cost function. Although is has been shown that clever weight settings of OSPF doesn t always perform within a few percent of General Optimal Routing. But by adjusting OSPF weights we can increase the networks ability to honor increasing network demands. The optimization of Project Final-Report Topics in Internet Research (CS 678) Page 23 of 23

24 the OSPF weights in this way is able to produce results that are comparable with more flexible routing schemes like MPLS etc. 5.4 Weight Management by Some External Entity In another approach [19] authors have described a methodology to engineering the flow of traffic in IGP networks. It is shown that by monitoring the traffic and topology, optimizing the setting of the static link weights, and reconfiguring the routers with new weight settings as needed leads to a comparable performance gain; almost comparable to the General Optimal Routing. The above mentioned approach treats traffic engineering as a network operations task; something that is to be handled externally from the routing protocol implemented by the routers. The link weights are configured by an external entity, something like a network management tool or a human operator. The modification of the link weights is considered a significant change to the underlying network and is performed infrequently. The centralized (network wide) approach taken by the authors for setting routing weights has some benefits like protocol stability because of non-frequent changes in weights unless the network topology changes or parameters are reconfigured. Another benefit is low protocol overhead as the routers need not to keep track of the changes in a link s load and flood the information to the network The Centralized Approach The centralized approach to traffic engineering has following main steps to perform on the underlying IP network:- 1. Measure the network topology and the offered traffic on the network. 2. Model the underlying network by evaluating possible weight settings of the link weights and predict how the IGP configuration will affect the traffic flows. Project Final-Report Topics in Internet Research (CS 678) Page 24 of 24

25 3. Control the optimization of the underlying network by changing the IGP configuration on one or more routers. The link values that are to be deployed are the ones calculated in the previous step. Topological or configuration information about the underlying network can be obtained from a number of sources like some parameters in Link State Advertisements (LSA) of the routers, Simple Network Management Protocol (SNMP) etc. An integrated network management system automates the entire process of detecting congestion, re-selecting (re-computing) suitable IGP weights and effecting / changing the network configuration. After a weight change by integrated network management system, the router updates its link-state database and floods the new value of the weight to the rest of the routers in the network. Upon receipt of the new link-state advertisement, each router updates its link-state database, re-computes the new shortest paths, and updates the relevant entries in its forwarding database. During all this phase the routers in the network may not have a consistent view of the shortest-path routes to some destinations Approaches to Assigning Link Weights A comparison is made between different approaches to assign a weight to a link [3]. Three approaches are compared; UnitOSPF (a cost of one is assigned to every link), InvCapOSPF (weight is inversely proportional to the link capacity) and AdvancedOSPF (mentioned above as piece wise linear cost function or in [17]). The conclusion is that the AdvancedOSPF performs within a few percent of the OPT and other two perform quite poorly. AdvancedOSPF is able to handle 70% more demands as compared to the other two approaches to assign weights to links. It is important that the IGP weight settings be robust to the changes in the traffic and the underlying network topology. Although limiting the number of weight changes is important to limit the disruption to the network but changes to the link weights are necessary in response to large shifts in the traffic or a router / link failure. The conclusion Project Final-Report Topics in Internet Research (CS 678) Page 25 of 25

26 is that despite the advantages of the more advanced routing protocols, the simple approach of adjusting the static link weights still remains practicable for many IP networks. 5.5 Dynamic Optimization Using Adaptive Random Search In [19] authors have proposed an improvement upon the static local search heuristic approach (discussed earlier). Authors have also solved some of the problems linked with the dynamic optimization of OSPF weights. GI/M/1/K queuing model is used to analytically compute the total packet drop rate in the network. Dynamic optimization problem was formulated by choosing the total packet drop rate as the optimization criteria. Authors in [19] have demonstrated substantial improvement in performance in terms of the packet loss rate by dynamic optimization of OSPF weights. A faster search scheme (Adaptive Random Search) as compared to the Shortest Path First (SPF) is proposed, which gives substantial improvement in the speed of optimization as well In this paper, authors consider a dynamic optimization problem based upon the fact that periodic information about the traffic demands is available. Periodic information about the traffic can be obtained by either monitoring the traffic at the edge routers or by the information provided by the SLA. The information about the traffic demands is then used to obtain the optimal link weight setting for the current traffic requirements. Following components are needed for the dynamic optimization of OSPF weights:- 1. An automatic network management tool. The tool is required to monitor the traffic to obtain the demand estimates and to deploy the estimated optimal link weight settings in the network. 2. A method to search large parameter spaces to obtain link weight settings for the performance improvement. Local search schemes cannot be used in this case because of inefficiency in high-dimensional optimization Project Final-Report Topics in Internet Research (CS 678) Page 26 of 26

27 problem and susceptibility to noise in the objective function. Hence, the local search scheme proposed in [17] and discussed earlier cannot be used for dynamic optimization. The deployment of the Improved OSPF weights in the real network dynamically is achieved by using the Online Simulation (OLS) framework. OLS is a general automatic network management framework that tunes the parameters of the network protocol Packet Drop Probability Higher drop probability is observed when the arrival process has a high variance i.e. when the incoming traffic is bursty. P l = (t l p l ) (p l +1 t l ) K l / 1 (p l +1 t l ) K l +1 Where: t l = a X / B l p l = 2λ l / σ l λ l a l = p l λ l In short, the above equation [19] represents the expression of packet drop probability, P l on a single link l as a function of mean λ l, variance σ 2 of the arrival process, mean packet size X, link bandwidth B l and buffer space K l. For further explanation of the parameters and the equation please refer to [19] The Adaptive Random Search Algorithm The goal of the proposed search scheme [19] is not to search for the optimum parameter setting, but to find a better parameter setting within a given amount of time. Another requirement is that the search algorithm must be able to handle a large number of parameters. Project Final-Report Topics in Internet Research (CS 678) Page 27 of 27

28 The algorithm does not use any traditional local search method. It is more robust to noises in function evaluations and is scalable to high dimensional problems. The algorithm first does a random sampling of the entire parameter space and identifies a promising region. After this the algorithm starts another random sampling in the promising region. Using this methodology the sample space is shrunk and realigned based on samples in the promising region only. The whole process continues until it finally converges to a local optimum. The process is repeated for the whole network until the stopping criterion is satisfied. Some simulations are done on ARPANET, MCI and randomly generated network topologies [19]. Simulations show that the adaptive random search scheme out-performs the local search scheme in terms of the number of iterations needed to find a good parameter setting for all the three given network topologies. The adaptive random search algorithm took 50-90% less iteration as compared to the local search algorithm given in [17] and discussed above. It is also shown that for the packet drop rate metric, the adaptive random search scheme takes 50% or less iteration to figure out a good link weight setting. An overall improvement of % is obtained in the total packet loss rate in the network. 5.6 Integrated Approach to Traffic Engineering Authors in [18] propose a new approach, called the integrated approach, to achieve traffic engineering without using any overlaying. Traffic engineering objectives like load balancing etc are achieved through manipulating link metrics for IP routing protocols (such as OSPF or IS-IS). The underlying idea of the approach is that the optimal routing with respect to any objective function, is always shortest path routing with respect some appropriate link weights settings. Project Final-Report Topics in Internet Research (CS 678) Page 28 of 28

29 Authors have used a systematic method for deriving the link metrics. This method is used to convert a set of optimal routes for traffic demands to shortest-paths with respect to the link weights. Later it is shown that this integrated approach can achieve exactly the same results as any other overlay approach. Advantages of this approach include simplicity of IP routing and little changes to the Internet architecture. Once the link weights are calculated and deployed, the shortest-path routing protocol such as OSPF can be used to calculate the shorted paths in term demand of the underlying network in the usual way. The proposed approach is not restricted to any particular objective function like minimizing congestion on a link of maximizing throughput. The methodology can be applied to any of the objectives as required The Objective function Authors in [18] have modeled the IP network as a digraph G = (V, E) where V is the set of nodes and E is the link between them. The links have fixed capacities, and the traffic demands between the nodes are also given. The links and their capacities are directional, i.e. link (i, j) is considered different from link (j,i) and same goes for the bandwidth of the link. K is the set of point to point demands. For each k that belongs to K d k, s k, t k represent the size of the demand bandwidth, the source node, and the destination node respectively. w ij is a given set of link weights where i,j belong to E. The Variable X k ij represent the percentage of demand k which flows across the link (i, j). The objective function is follows:- min s t Σ k є K Σ (i, j) є E ( w ij X k ij ) The above objective function is used to minimize the total weight of the links used from source s to destination t. Project Final-Report Topics in Internet Research (CS 678) Page 29 of 29

30 Authors further show that for any arbitrary set of routes, as long as it is not loopy, can be converted to shortest paths. And for any loopy set of routes, there is always a better set of routes that consumes less bandwidth the shortest path. (A set of routes R1 is called loopy as defined by authors in [18] if there exists another set of routes R2 such that on every link, R2 uses bandwidth less than or equal to R1 and one at least one of the links R2 uses bandwidth strictly less than that used by R1). 5.7 Enhancement to Static Local Search Heuristic Approach Authors in [21] say that two major assumptions made in static local search heuristic approach (discussed earlier) to traffic engineering are at odd with the IP forwarding mechanisms. Assumptions are as follows:- 1. Forwarding decisions are specific to source destination pair 2. Traffic can be split in any arbitrary ratio over shortest paths In [21] authors say that even splitting of traffic across multiple shortest paths is a more challenging task. In their proposed approach instead of changing the forwarding mechanism responsible for distributing traffic across equal cost paths, a set (or a subset) of shortest paths (next hops) is assigned to routing prefix entries in the forwarding table(s) of a router. This allows controlling traffic distribution without modifying existing routing protocols (like OSPF and IS-IS) and forwarding mechanisms. Detailed discussion of this approach is beyond the scope of this paper. For details please see [21]. Project Final-Report Topics in Internet Research (CS 678) Page 30 of 30

31 6 Conclusion In this report first Traffic engineering is looked upon as a fundamental technique to enable high performance networks utilization and management as the intoxication of the Internet and Intranets is growing day by day. After that some congestion control mechanisms are discussed since it is one of the major performance bottlenecks. The main focus of the paper was on Intra-domain traffic engineering. How traffic can be engineered within a given AS. Two main approaches to Intra-domain Traffic Engineering are used as discussed in the paper. One approach is to try and optimize the exiting route forwarding methodology and the other approach is to use some emerging constraint based route forwarding methodology like MPLS. It has been shown in the paper that it is possible to cleverly adjust the link weights of the network to make it perform with a few percent of the optimum deployment of traffic on the underlying network. The end result is that IGP Traffic Engineering methodologies can still make underlying network perform almost as good as would have achieved by deploying any other emerging methodology like MPLS. Therefore the switch from currently used IGP route forwarding mechanisms to other emerging mechanisms like MPLS may not be required. Project Final-Report Topics in Internet Research (CS 678) Page 31 of 31

32 7 References [1] D. O. Awduche, A. Chiu, A. Elwalid, I. Widjaja, and X. Xiao, Overview and principles of Internet traffic engineering. Request for Comments 3272, May [2] B. Olivier, Internet traffic engineering techniques, URL: [3] B. Fortz, J. Rexford and M. Thorup, Traffic Engineering With Traditional IP Routing Protocols, IEEE Communication Magazine, October [4] W.T. Zaumen and J.J. Garcia-Luna-Aceves. Dynamics of Distributed Shortest- Path Routing Algorithms. Computer Communication Review, 21(4):31ñ42, September Proceedings of ACM SIGCOMM (p 27) [5] Nichols, K., Blake, S., Baker, F. and D. Black, "Definition of the Differentiated Services Field (DS Field) in the IPv4 and IPv6 Headers", RFC 2474, December [6] Moy, J., "OSPF Version 2", STD 54, RFC 2328, July [7] Castineyra, I., Chiappa, N. and M. Steenstrup, "The Nimrod Routing Architecture", RFC 1992, August [8] D. Awduche, "MPLS and Traffic Engineering in IP Networks", IEEE Communications Magazine, Dec [9] Rosen, E., Viswanathan, A. and R. Callon, "Multiprotocol Label Switching Architecture", RFC 3031, January [10] Shenker, S., Partridge, C. and R. Guerin, "Specification of Guaranteed Quality of Service", RFC 2212, September Project Final-Report Topics in Internet Research (CS 678) Page 32 of 32

33 [11] Wroclawski, J., "Specification of the Controlled-Load Network Element Service", RFC 2211, September [12] Bernet, Y., Ford, P., Yavatkar, R., Baker, F., Zhang, L., Speer, M., Braden, R., Davie, B., Wroclawski, J. and E.Felstaine, "A Framework for Integrated Services Operation over Diffserv Networks", RFC 2998, November [13] Yavatkar, R., Pendarakis, D. and R. Guerin, "A Framework for Policy-based Admission Control", RFC 2753, January [14] K.M. Chandy and J. Misra. Distributed Computationon Graphs: Shortest Path Algorithms. Communications of the ACM, 25(11):833ñ837, November (p 26) [15] E.W. Dijkstra. A Note on Two Problems in Connexion with Graphs. Numerische Mathematik, 1:269ñ271, (p 26) [16] R. Perlman. Interconnections. Addison Wesley Longman, 2nd edition, (pp 26, 27) [17] B. Fortz and M. Thorup, Internet traf_c engineering by optimizing OSPF weights, in Proc. IEEE INFOCOM, March [18] Z. Wang, Y. Wang, and L. Zhang, Internet traffic engineering without full mesh overlaying, in Proceedings of INFOCOM 2001, Anchorage, Alaska, April [19] T. Ye, H. Kaur, S. Kalyanaraman, K. Vastola and S. Yadav, Dynimiac optimization of OSPF weights using online simulation, in Proc. IEEE INFOCOM, March [20] R. Coltun, The OSPF Opaque LSA Option. Request for Comments 2370, July 98. [21] A. Sridharan, R. Gu erin, C. Diot, Achieving Near-Optimal Traffic Engineering Solutions for Current OSPF/IS-IS Networks IEEE, Project Final-Report Topics in Internet Research (CS 678) Page 33 of 33