The Internet now-a-days has become indispensable to each and everyone. It is vulnerable to

CHAPTER 5 NETWORK AVAILABILITY IN INTERDOMAIN ROUTING The Internet now-a-days has become indispensable to each and everyone. It is vulnerable to node failures, link failures, and fluctuations due to many known or unknown reasons in the network connectivity. The bitter truth is, even today networks' failure, link faults are happening. A single change in a link or a node has a potential to trigger the unstablerouting-tables of many nodes. These failures may lead the network to an unstable state by increasing its convergence time significantly longer. In this chapter an algorithm is proposed to keep the value of the minimum route advertisement interval (MRAI) timer variable unlike the conventional approach of keeping it constant. The proposed approach makes the timer value varying depending on the network conditions with respect to the origin of the prefix advertised on the network. Simulation results show that the convergence time becomes significantly low and helps the network converge quicker [Kumar and Kumar (2013b)]. 5.1 INTRODUCTION T he Internet experiences changes in connectivity due to rapid changes in topologies. The symptoms of these route instability cause non-availability of those routes which previously existed in the routing table. Whenever a route is withdrawn from the router s routing table it selects the new route if available and advertises the newly selected route to its peers otherwise it sends a withdrawal of that route to its peers. But sometimes this simple looking process lasts for the relatively longer duration and it ends with the exchange of many

updates. Therefore the network may remain unstable during this period and convergence is not achieved immediately, rather it takes relatively longer time. Sam Halabi [Halabi (2000)] has identified some factors that affect network connectivity, as follows: Faulty hardware Software problems Insufficient CPU power Insufficient memory Network upgrades and routine maintenance Human error Link congestion Interior Gateway Protocol instability The period for which the BGP convergence process lasts is referred as the BGP convergence time. When a route to a destination prefix is changed (advertisement or withdrawal ), it may lead to the exchange of a series of messages. The convergence time is defined as the time elapsed from the instant when the first update message containing a change of the destination reachability is sent, and the last update messages, related to this change, are received [Griffin and Premore (2001)]. To control these types of instabilities during the convergence time, initially BGP designers did not make any provision. But later the provision was done by the introduction of a timer called minimum route advertisement interval timer, which is popularly known as the MRAI timer. This MRAI timer is used to control the update sending frequency for a particular peer about a particular destination. This timer delays the updates for a time either decided by the implementer or by default its value is set to 30s. The internet is a collection of various networks managed by Network Providers, Internet Service Providers, Universities, Organizations, Private Companies etc. These networks are heterogeneous in nature. Each of these networks, managed by an administrative team, is called an Autonomous System (AS). Before reaching final destination traffic is forwarded through many of those ASes. The network structure varies according to its locations. That means the network connecting end user, internet service providers (ISPs), and universities are different. To communicate with different organizations or service providers some routers are designated as 95 P a g e

border speaker routers by each provider/organization. The interconnection between border speaker routers, from different organizations, is popularly known as interas network. The de facto protocol for this purpose is Border Gateways Protocol (BGP). The process of routing between ASes is also called an interdomain routing [Halabi (1997)]. The BGP makes a network of ASes for exchanging information with immediate neighbors. The network formed by BGP may be seen as a tree because it sees the entire internet as a graph of ASes. A unique number is used to identify each AS called AS number (ASN). The BGP prepares a list of ASNs appearing from source to destination called a path as shown in Figure 5.1. It uses a path vector algorithm to compute the best path among multiple available paths from a source to a destination. The path information associated with destination prefix is used to ensure that path is free from loops. Figure 5.1: InterAS Network A router that runs a BGP routing process is called the BGP speaker. The BGP speakers, when establishes a peering connection with each other to exchange routing information, are called peers or neighbors. Transport layer protocol, TCP port 179, is used to communicate with these peers to make the communication between peers a reliable. Therefore it simplifies the complexity related to reliable design [Rekhter and Li (1995 )]. 96 P a g e

In the beginning the BGP speakers exchange all BGP routes with their peers after the session is established between them. Once the session establishment is completed and initial routes have been exchanged with the peers after that only incremental route updates are exchanged, this approach of exchanging only incremental updates proved comparatively beneficial in terms of computational overheads and bandwidth [Cisco (2012)]. The BGP speaker exchanges its routes with its peer in the form of update messages. The update message contains prefix information, path length, and path attributes. The prefix information, and path length contains the information for the neighbor about all the routers in the way of reaching the destination. The path attributes include information such as the list of ASes that are to be traversed and the degree of preference for a particular path. Due to any reason if a network prefix becomes unreachable that path is called an invalid path, and information of invalid paths are propagated to peers in the form of withdrawal messages that is part of the update messages [Cisco (2012)]. Paths that are no longer available or information associated with the path is changed or a new path for the same prefix is selected then a withdrawal message is not required, in that case a replacement path is advertised. A BGP speaker, to check if the connection with the neighbors is alive, periodically sent keepalive messages. The keepalive messages require very small bandwidth so it is not considered as a burden either on the CPU or on the bandwidth. Hold timer may be used as an indication of the time elapsed between successive update/keepalive messages exchanged [Labovitz (1998)]. To keep track of the current instance of the BGP routing table a table version number is used. Whenever any BGP router notices a change in its routing table immediately it increments its table version number. Therefore if the table version number increases at a rapid rate it may be inferred to network instability. To control the network instability route flap dampening and other provisions have been designed. Different stages of the negotiation process are required for connection establishment, as shown in figure 5.2 that shows major events with the help of open, keepalive, and notification messages sent to the peer from one stage to another [Griffin and Premore ( 2001 )]. 97 P a g e

Figure 5.2: BGP Negotiation Process [Cisco (2012)] Convergence time is the amount of time required for all the routers in the network to become updated with all the new routes, changed routes, or disappeared routes. Earlier, many studies have been shown to improve BGP convergence time by lowering down the default value of minimum route advertisement interval timer, classifying and delaying the update messages, limiting the withdrawal message frequency etc. But none of them have shown the variation in timer depending on the changing network conditions. 5.1.1. ROUTE PROCESSING IN BGP There are two types of peers in BGP, a first type of peers are those peers that are from the same Autonomous System, and second type of peers are from different Autonomous Systems. Those peers that are from the same AS are called internal peers; the connection is established with 98 P a g e

ibgp session. And those peers that are from different AS are called external peers; the connection is established with ebgp sessions [Dan Pei et al. (2005)]. The BGP is a path vector protocol; it sends full path information to the peers so that if there any loop it can be identified and the path with the loop can be avoided. After receiving updates it computes the best path among all the available paths to the destination, and then that best path is installed in its forwarding table, and updates are sent out to the peers. Whenever there is any change in the routing tables because of new announcements, change of routes in preexisting destinations, or withdrawal of the route, table version number is incremented. In converged state of the network all the nodes must have the same table version number. C Labovitz et al. [Labovitz et al. (2000)] have proposed a method to improve the BGP s convergence by including additional synchronization, and diffusing updates in the protocol but all the proposed changes increases complexity and also increases router overhead. 5.2 ANALYSIS OF PREVIOUS WORK To address the problem of slow convergence in the BGP network several efforts have been made in the literature and the majority of those efforts emphasizes on keeping the BGP design unmodified and making temporary changes to the protocol. 5.2.1 BGP WITH AN ADAPTIVE MINIMAL ROUTE ADVERTISEMENT INTERVAL The author Laskovic [Laskovic (2006)] in his work has underlined the importance of the Minimal Route Advertisement Interval (MRAI) timer. The author is not in favor of keeping the MRAI timer value to its default level of 30 seconds, but it gives a method to minimize the values from 30s to the timer predicted based on previous value. Previous studies have also reported the existence of optimal MRAI values that minimize the BGP convergence time. The author has named his approach as the adaptive MRAI algorithm for dynamic MRAI timer values. The reusable MRAI timers have been used to control the number of advertisements for each destination. 99 P a g e

Figure 5.3: Intradomain and Interdomain Routing The BGP speaker routers interact with intradomain routing protocols which take the responsibility of decision making for user data forwarding within the Autonomous System, like Routing Information Protocol (RIP), Intermediate System-to-Intermediate System (IS-IS), or more popularly the Open Shortest Path First (OSPF), as shown in figure 5.3. The MRAI timer is implemented in two different manners, first with the BGP speaker routers, second on the routers within an Autonomous System. 100 P a g e

Table 5.1: Procedure for Adaptive Timer [Laskovic (2006)] Adaptive Timer Steps Initialize activen(d) = adaptivemrain(d) - idlen(d) avg_ activen(d) = f(avg_ activen-1(d), roundn(d)) deviationn(d) = f(deviationn-1(d), roundn(d)) IF idlen(d) < 1 s Then adaptivemrain+1(d) = 2 x avg_ activen(d) ELSE adaptivemrain+1(d) = avg_ activen(d) + 3 x deviationn(d) IF adaptivemrain+1(d) > 30s Then adaptivemrain+1(d) = 30 s ELSE reusable_timern+1(d) = te + adaptivemrain(d) idlen+1(d) = 0 ; roundn+1(d) 101 P a g e

Figure 5.4: (a) Convergence time, (b) updates exchanged The performance of the author s approach for keeping the MRAI timer value smaller results in better network convergence time, and but the number of updates exchanged in this process are more than what it was with default relatively large value of the timer. Therefore it seems an improvement in terms of getting smaller network convergence time while on the other front the overheads are high because more the messages exchanged incur more bandwidth consumption and more computing / processing power. 102 P a g e

5.2.2 IMPROVED BGP CONVERGENCE VIA GHOST FLUSHING The authors Yehuda Afek, Anat Bremler-Barr, and Shemer Schwarz have proposed through their work an approach to improve the convergence time of the network of the BGP speakers. They have proposed an algorithm which enables the BGP speakers to inform their peers when any route is no more available for use. The authors have named the advertisements containing invalid routes as a ghost information and have given an algorithm they called it the ghost flushing (GF) algorithm to remove it from the routing table of routers [ Afek et al. (2003)]. The BGP speakers first advertise their reachability information to all of its peer speaker routers, and then each router selects the best path and installs it in the routing table. Whenever any BGP speaker changes its known best routes to a destination, it has to inform its peers by sending an update about the change of the best path. However, sending updates may get delayed because of the MRAI timers. If the updates are delayed, then the peers keep the invalid routes in their routing table that may be propagated further. To avoid the delay, the authors have the mechanism which called ghost flushing (GF), it sends withdrawals immediately to its peers, and informs them that a route is not available. These withdrawals are called by the authors as flush withdrawals and these withdrawals format is as defined in RFC 1771 [Rekhter and Li (1995]. Table 5.2: Algorithm GF [ Afek et al. (2003)] Algorithm If (NewASpath dst less preferred than LastAnnouncedASpath dst ) {An empty path ({}) is considered longer than any other path} If (currenttimestamp LastAnnouncedTime dst < minrouteadver) Send message (withdrawal,{},dst) to each peer LastAnnouncedASpath dst = {} 103 P a g e

The operation of the algorithm has been shown in figure 5.5(a) - (d), which shows the flushing of ghost information from all the nodes. Figure 5.5: Flushing Ghost Information Normally the withdrawals are sent to inform a BGP router that it does not have any route to the particular destination, but the withdrawals in the proposed approach are used to remove an invalid route. When the MRAI timers expire, the BGP speaker send the new best route and the peers will update their routing tables. The ghost flushing generates a storm of withdrawals that removes invalid routes. The proposed approach has several limitations [Mao et al. (2002)]. The first limitation is that it may not eliminate all invalid routes, because of different propagation delays. The second limitation is that the proposed method may induce hike in number of update messages exchanged, which may result in route suppression because of route flap damping mechanisms [Wang et al. (2002), Mao et al. (2002)]. The third limitation is that it requires extra cost in terms of additional memory for storing the last update message sent to each peer. 104 P a g e

5.2.3 DIFFERENTIATED BGP UPDATE PROCESSING FOR IMPROVED ROUTING CONVERGENCE The Authors Wei Sun, Zhuoqing Morley Mao, and Kang G. Shin have proposed differentiation of updates depending on the priorities, which are based on whether it lies on the route, off the route, or it belongs to the provider or the customer. If they are lying on the route or it belongs to customer, they are processed with higher priority. There are two issues in the proposed method [Sun et al. (2006)]. First point is locally inferred routing preference: As per this preference, when a router is to send updates to a particular peer, it checks if that peer router has sent updates for the same destination prefixes to itself. If yes, then it is categorized to be of low priority. Second point is difference based route selection: In this issue, when failure happens, the best route to a destination prefix is withdrawn, instead of selecting the next best route (if available), a router first selects an temporary route which is different from the withdrawn route. Figure 5.6: Update Division in Priority 105 P a g e

Table 5.3: DUP-Algorithm Algorithm // AS1 sends an update for D to AS2: IF ASP 2 = Null, Then sends the update with shorter MRAI timer; Else IF AS1 is a peer or provider of AS2, Then sends the update with longer MRAI timer; Else // AS1 is a customer of AS2 IF ASPpre1 = NULL, Then sends the update with shorter MRAI timer; Else IF len (ASP pre1 ) len (ASPnew1 ) IF len (ASPnew 1 ) + 2 < len (ASP2), Then sends the update with shorter MRAI timer; Else sends the update with longer MRAI timer; The proposed DUP algorithm claims to reduce the number of low-priority updates and the routing convergence time. The authors claim to improve the network availability. 5.3 ISSUE OF INTRERDOMAIN ROUTING A long convergence time has been a critical issue for any network provider or service provider, because the long convergence time affects the data forwarding process for a significantly longer duration and till that time network remains unavailable. 106 P a g e

5.3.1 MRAI TIMER VALUE AND CONVERGENCE TIME IDENTIFICATION Long convergence time on any network has negative impact on its reliability because of the compromised packet delivery. Whenever any change in the network topologies is caused either due to the introduction of new network prefixes announced in any update or because of the withdrawal of previously existing network prefixes due some known or unknown causes of withdrawals. Many previous works have highlighted the importance of the Minimum Route Advertisement Interval (MRAI) timer in network convergence. The default value of MRAI timer is 30 seconds in many products coming from various vendors. But this default value is the main cause of long convergence time hence the network remains unstable for longer duration therefore affecting packet delivery. 5.3.2 MRAI TIMER, CONVERGENCE TIME, AND NETWORK TOPOLOGY The analysis of convergence time on different topological structures shows that the convergence time not only depends on the MRAI timer but also on the varying size and type of network topology in which all the nodes are connected to each other. The Analysis results are shown with the help of figures 5.7, 5.8, and 5.9. From the aforementioned figures one can analyze that initial connectivity is delayed in all the layouts of connecting nodes together but some of the layouts have significantly large delay which makes them less preferable in comparison to others. As it can be seen from figures 5.7, 5.8 and 5.9 that the linear layout of connecting nodes is the least preferable and mesh layout which connects every node with every other node directly seems more preferable but from the network`s point of view it is not tough to understand that this completely connected layout, which is mesh, is not practically feasible. So the next choice left is the combination of more than one layout together named as random in the aforementioned figures. 107 P a g e

Delay (in seconds) Topological layouts were simulated in Qualnet 5.2 [SNT (2012)] simulator in a windows 7 professional environment. The size of the topologies varied from 4 nodes to 50 nodes and obtained results were analyzed. 70 60 50 40 30 20 10 Connectivity Delayed 0 Ring Line Mess Random Type of Topology (4 ) Figure 5.7: Delay in Initial Forwarding in Four Nodes Topology Figure 5.7 above shows the results for the four different topological layouts with the total number of nodes in each topology were only four. The initial forwarding delay is the time required for all BGP routers to get complete consistent forwarding information required to forward user data. As figure 5.7 shows the linear topology where all the nodes were arranged in a line and connected in sequential manner has longest delay. The results obtained by increasing the number of nodes in each topology explain that increment in delay is multi folded. 108 P a g e

Delay (in Seconds) Delay (in Seconds) 300 Connectivity Delayed 250 200 150 100 50 0 Ring Line Mess Random Type of Topology (10) Figure 5.8: Delay in Initial Forwarding in Ten Nodes Topology 700 600 500 400 300 200 100 Connectivity Delayed 0 Ring Line Mess Random Type of Topology (20) Figure 5.9: Delay in Initial Forwarding in Twenty Nodes Topology 5.4 PROPOSAL FOR VARYING MRAI TIMER VALUE The analysis of MRAI timer is for its optimum value, the value of the MRAI timer is variable and it will be dependent on the two factors; first is the cumulative delay of the links connecting the destination, second factor is the maximum computation delay of the heavily loaded node on 109 P a g e

path of the network. These attributes are to take care of the dynamic delay and the load onto the network at the time of disturbance. The length of the paths available with a node in its routing table is considered here because the different paths may have different delays, therefore it s important here. But whenever the contributing factors sum exceeds default value of the MRAI timer which is 30s, the default value of 30s will be used. Here the objective is not to suppress every update to every peer for 30 seconds unnecessary if it can work fine with lower value and can also save the network time. Therefore the update suppression may vary up to a maximum of default value. Let P be the set of paths that lead to the destination D from the current node. Pi P, Pi is the i th path to the destination D P i = { r1 r2.. rn} Where n N, which is the number of routers on the path P. Where r i R R is the set of routers lying on the path to destination D. R = { r1,r2,r3...rn } L(Pi) is the set of links lying on the path to D. L(Pi) = i R(D) { li } R(D) is set of router on the path to D. W(L(Pi)) = i R(D) { W ( li ) } Where w(li) is the delay of the link li. P(w)= max { (w (L(Pi)) } Where P(w) is the maximum weight of the path to destination D. Q is the computation delay of the router on the paths to D. 110 P a g e

Q(R)= max rϵr { Q(r) } mrai = P(w) + Q (R) Table 5.4 VBGP Algorithm ALGORITHM VBGP 1. BEGIN 2. Check prefix entries in the routing table 3. Get the paths to the destination D and stored in P 4. Get the delay P(d ) for each link on the AS path from S to D store in Path 5. Get the processing load P(R) of each router in AS Paths 6. Get the number of routers in N 7. For each path in P do 7.1 Path Delay max {P(d)1, P(d)2, } 8. For each router in R do 8.1 Processing Delay max { P( R )1, P ( R2 ) } End For 9. MRAI Path Delay + Processing Delay 10. IF MRAI gt 30 Then 10.1 MRAI Default 10.2 Return MRAI End IF 11. ELSE Then 11.1 Return MRAI End ELSE 12. END 111 P a g e

The simulation is performed on the topology shown in figure 5.10. In figure 5.10 each node represents a BGP router which is a BGP speaker for its domain and also actively participates in path calculation, timer computation, and in forwarding updates. Simulation was performed using network simulator ns-bgp [NS2 (2008)]. The BGP implementation is based on BGP-4 specifications. The network topology is manually created, each node represents an AS, and all the links are configured to have similar bandwidth and different delays. Each node advertises some prefixes to all its peers. The nodes are having different load hence have different processing time for an update, we have varied the processing delay from 1ms to 200 ms. The link delays are also different and varies from 1ms to 1000 ms. The node 0 announces a prefix, and after small delay it withdraws the earler announced prefix. The observations are done at node 6. The mrai timer computed and its maximum value was recorded to be 2.600s. There is significant improvement is initial network up time. Figure 5.10: Simulation Topology Figures 5.11 5.12 depict how the convergence time varies at the different values of MRAI timer so as the number of updates exchanged between peers. The convergence time increases with the increment in MRAI timer value, and number of updates decrease as the MRAI timer 112 P a g e

Number of Updates Convergence Time (Seconds) value increases. Therefore it is easy to understand the contradiction in convergence time and the number of updates exchanged. 100 80 60 40 20 Network Convergence Time 0 5 10 15 20 25 30 MRAI (in Seconds) Figure 5.11: Convergence Time of Line Topology for Different MRAI Timer Values Impact of MRAI on Updates 40 39 38 37 36 35 34 33 32 31 5 10 15 20 25 30 MRAI Updates Figure 5.12: Impact of MRAI Timer on Updates After analyzing figures 5.11 5.12, it is proposed to keep the MRAI timer value less than its default value wherever it is possible to keep the convergence time relatively low. The proposed algorithm VBGP proposes to keep the value of the MRAI timer depends on the path delay and the processing delay. It is to be maintained on per peer per prefix basis. The algorithm implementation results are shown in figures 5.13 5.15. In these figures the results show that the 113 P a g e

Number of Updates Number of Updates convergence time is significantly lower, the number of updates are more. When we analyze the performance of the approach during link flaps the results are, shown in figure 5.14-5.15, better in terms of bringing the network to a converged state than default MRAI implementation, and marginally better than BGP-AM [Laskovic and Trajkovic (2006 )]. 1200 Number of Updates 1000 800 600 400 200 BGP VBGP BGP-AM WRATE+SSLD 0 10 20 30 40 Number of Nodes Figure 5.13: Comparative Analysis of Updates in Growing Networks Size 160 140 120 100 80 60 40 20 0 Updates Exchanged During Flaps 1 4 10 BGP VBGP BGP-AM Figure 5.14: Comparative Analysis of Flaps on Updates 114 P a g e

Time(Seconds) 35 Network Recovering During Flaps 30 25 20 15 10 BGP VBGP BGP-AM 5 0 1 4 10 Figure 5.15: Comparative Analysis of Flaps on Convergence Time Similarly when we compare the number of updates exchanged during flaps, the result shows, in figure 5.14, close to the BGP-AM, but not with the default BGP implementation. As the computation power increasing every year, these additional updates will not have significant impact of the performance. Therefore the performance of the approach VBGP is comparable and benefits recovering the network faster, which yields better packet delivery and low loss of user data during troublous times. 5.5 CONCLUSIONS The unprecedented growth of the internet has put pressure on the designers to make the network reacting quickly to each and every, small or big change in the connectivity and overcome the problem and bring the whole network available in minimum possible time. An attempt made through this work, by introducing an algorithm to keep the MRAI timer value based on present network conditions, to enable the network to react promptly to the changes occurred in the network connectivity. The results verify the usefulness of the proposed approach in bringing the network convergence time low. Therefore the network quickly becomes available for use. 115 P a g e