Adaptive Load Balancing based on IP Fast Reroute to Avoid Congestion Hot-spots

Transcription

1 Aaptive Loa Balancing base on IP Fast Reroute to Avoi Congestion Hot-spots Masaki Hara an Takuya Yoshihiro Faculty of Systems Engineering, Wakayama University 930 Sakaeani, Wakayama, , Japan Abstract Several loa balancing techniques for IP routing scheme have appeare in the literature. However, they require optimization process to compute optimal paths to meet traffic eman so that it requires a mechanism to measure traffic eman an to share them among all routes in orer to follow ynamics of traffic. It naturally results in communication overhea an losing sensitivity to follow traffic ynamics. In this paper, we investigate a loa balancing mechanism from another approach, i.e., base on IP fast reroute mechanisms. The main iea is simply to forwar packets into etour paths supplie by IP fast reroute mechanisms only when packets meet congestion. This strategy enables us to use vacant resources aaptively as soon as they are require to avoi an issolve the congestion. Through traffic simulation we show that IP fast reroute base loa balancing mechanisms improve the capacity of networks. I. INTRODUCTION Recent growth of the Internet an the rapi increase of user traffic require more banwith an communication quality for infrastructure networks. To make the most of existing network resources, several traffic engineering (TE) techniques have been propose so far. Currently TE techniques for MPLS scheme have become important [1] an several techniques are working in practice mainly in backbone networks. However, since MPLS capable routers are expensive, IP routing is still use in consierable part of networks. For IP routing scheme, several TE techniques have also been propose. Forts et al. [2] first presente the metho to optimize link metrics. They formulate the link metric optimization problem by means of linear programming which compute the optimal link metrics to make the most of network resources assuming that traffic eman among every pair of noes is given. Their work also presente that IP base TE achieves the similar level of loa balance performance compare to MPLS base TE techniques. However, it is consiere true that, in IP base shortest-path scheme, traffic tens to be concentrate on particular links or noes since the shortest-paths ten to use the common low-cost links. On this point of view, Dasgupta et al. [3] presente a result of performance comparison between the metric-optimize IP shortest path scheme an the MPLS base traffic engineering scheme. Their simulation emonstrate that uner the traffic variation coming from link failure, IP base TE often brings severe congestion hotspots cause of concentration of shortest paths, while MPLS base TE cause much less congestion. This result implies that there is still room to improve IP base TE techniques. As a more efficient loa balancing approach, several multipath loa-balancing schemes have been propose as IP base TE technique. Mishra et al. presente OSPF-base TE technique S-OSPF (Smart-OSPF) [4], in which source noes istribute traffic to their neighbor noes to try loa balancing among the shortest paths from neighbor noes. Antić et al. presente TPR (Two Phase Routing) [5] which once forwars packets to some intermeiate noes using IP tunnels an then forwars them to their estinations using normal shortest paths. They are capable of more efficient loa balancing than metric-optimization base traffic engineering. They both, however, require the optimization process which computes the paths for traffic istribution from traffic eman matrix. In practice, the optimization process requires the mechanisms to measure the traffic eman, to share the measure information among all noes, an to re-compute the new istribution paths. This process woul results in losing sensitivity for ynamic transition of traffic. In this paper, as another approach of loa balancing in IP routing schemes, we propose an IP fast reroute base loa balancing mechanisms. IP fast reroute is the technique to prevent packets from losing in case of link/noe or component failure using pre-compute alternative paths. IP fast reroute techniques achieve very fast recovery against failure, but its pre-compute paths are use only in case of failure. The loa balancing mechanism we propose utilizes those unuse paths even uner normal state to reuce congestion hot spots of networks. The approach of loa balancing using IP fast reroute has several goo characteristics as follows: 1. Packet forwaring is base on local ecision so that no aitional communication overhea (especially on traffic eman) is require for loa balancing. 2. No optimization process is require to follow ynamics of traffic so that sensitive response is achieve for traffic transition. This property is effective especially in case of failure protecte by IP fast reroute mechanisms. 3. Possible to work seamlessly with IP fast reroute mechanisms which provie failure protection function within the same framework. 4. Since only the packets face to be roppe are etoure, communication paths are basically the same as the shortest-path routing scheme. Namely, the current operational experience over traffic estimation is available /11/$ IEEE

2 x y estination x y (a) (b) Fig. 2. The Propose Loa Balancing Mechanisms x (c) y Fig. 1. The SBR Mechanism Primary tables Backup tables Backup tables with switch label as it is. Note that the iea of loa balancing using IP fast reroute is natural an may not be new. To the best of our knowlege, however, this work is the first specific mechanism escription an the first investigation on the performance of IP fast reroute base loa balancing mechanisms. The remainer of the paper is organize as follows: In Section 2 we escribe the base IP fast reroute mechanism calle SBR (Single Backup-table Rerouting) an present our loa balancing technique over SBR. In Section 3 we give the results of traffic simulation to show the performance of our mechanisms. In Section 4 we escribe several IP fast rerouting schemes propose ever an iscuss how we can implement loa balancing in them an the performance issue. Finally in Section 5 we give the concluing remarks. II. LOAD BALANCING MECHANISMS USING IP FAST REROUTE SCHEME SBR A. SBR Mechanisms In this section we explain the base IP fast reroute mechanism SBR (Single Backup-table Rerouting) [16] [17] an escribe how packets are forware into backup paths. SBR is an IP fast reroute scheme which recovers every single link failure with low overhea, i.e., two 1-bit flags in packet heaer an one aitional routing table. Note that there are several major IP fast reroute mechanisms stuie so far, but we use SBR as the base scheme since the mechanism is convenient to esign loa balancing functions. The consieration about other IP fast reroute scheme is seen in Sec. 4. SBR is first escribe in theoretical manner [16] an later its practical protocol to exten OSPF framework is presente in [16]. Note that the algorithm to compute backup routing table is similar to [19]; the main ifference as link protection framework is that SBR consiers link metrics to compute backup paths. The mechanism of SBR is illustrate in Fig. 1. Since we escribe SBR as an extension to the link-state routing scheme, it is assume that every noe has its primary routing table which represents the shortest-path tree for each estination. In SBR, every router has a seconary routing table, calle backup table, to recover any single link failure. Fig. 1(a) shows the situation that every noe (except ) in the network has its primary next-hop an backup next-hop for estination. Primary next-hops are inicate by soli arrows an Backup next-hops by two sort of broken arrows. Note that each backup next-hop (entry) is istinguishe into two types backup or switch, where switch type plays a special role in forwaring packets. See Fig. 1(b) for an example. Once a link (x, ) fails, the noe x etects it an immeiately forwars packets to u instea of, using its backup next-hop. Then the packets travel along backup next-hops for a while, an after going through that of switch type, i.e., (v,w) in this figure, they return to the primary route to reach the estination. Namely, packets sent from u estine to travel along the path u x u v w y. Such forwaring is realize by a 1-bit b-flag on packet heaer; keep b-flag 0 when a packet uses primary tables, change to 1 when the packet is firstly forware using backup tables (at x in Fig. 1(b)), then return to 0 when the packet passes a backup next-hop of type switch (at v in Fig. 1(b)) Namely, this flag represents which table to be use to forwar the packet. SBR is able to protect every single link failure. Fig. 1(c) is the example when link (y, ) faile. The packets sent from w estine to travel along the path w y v x. Aitionally in SBR, we introuce another 1-bit flag calle r-flag on packet heaer, which prevents loops in case of multiple link failure. The r-flag is set when a packet first returne to the shortest path (i.e., when the b-flag is first reset), an the packet with r-flag is never forware into backup nexthop. Namely, r-flag limits the number of etouring in orer to prevent loops cause of rerouting. B. Loa Balancing Mechanisms base on SBR Base on SBR, we propose a loa balancing technique to utilize vacant resources aaptively. Our strategy is simply to rescue the packets which are to be roppe as they have met congestion, by means of forwaring them into the backup path

3 TABLE I FORMAL FORWARDING PROCESS AT EACH ROUTER IN THE PROPOSED METHOD Flags an Counters (b,r) Conitions on Queue Process of Forwaring Flag Processing q (p) <T 1 forwaring into Primary Next-hop (0,0) q (p) T 1 an q (b) <T 1 forwaring into Backup Next-hop Increment r-counter, an set b-flag if the Backup Table entry is not s-labele. q (p) T 1 an q (b) T 1 rop q (p) <T 2 forwaring into Primary Next-hop (0,n) q (p) T 2 an q (b) <T 2 Backup Next-hop Increment r-counter, an set b-flag if the Backup Table entry is not s-labele. 1 n<max q (p) T 2 an q (b) T 2 rop (0,max) q (p) <T 2 forwaring into Primary Next-hop q (p) T 2 rop (1,n) q (b) <T 2 forwaring into Backup Next-hop Reset b-flag if Backup Table entry is s- labele. 1 n max q (b) T 2 rop instea of the primary path. Our metho etects congestion using output queue length. See Fig. 2. There is a router with three links. When a packet without b-flag arrive at the router, the router makes route ecision for it, i.e., ecies the output interface using the queue length for the primary next-hop of the packet. Namely, when the sum of queue length an the packet size is longer than threshol T 1 (T 1 is usually set as 100% of the queue length), the packet is enqueue into the queue of backup interface with the packet s b-flag set, an otherwise the packet is enqueue into the queue of primary interface. This simple rule, however, causes the etour chain problem. The etour chain problem is that etouring packets occurre from a congestion cause another congestion to generate etouring packets again, an this etouring continues like a chain. This problem must be avoie since the chain generates lots of etour packets which heavily consume network resources, an in the worst case they may form a loop, resulting in amplification of congestion. To prevent the problem from occurring, we introuce another threshol T 2 on the output queue to suppress etouring packets before they cause a new congestion. When a packet with b-flag or r-flag arrive at the router, the router checks the queue length of the backup interface to ecie the route to forwar it. If the queue length is longer than T 2 the router rops the etouring packet, an otherwise forwars it into the corresponing next-hop accoring to the forwaring rule of SBR. Note that we etour packets only if the packets are to be roppe uner the conventional single shortest-path scheme. Our strategy is to use vacant resources to rescue packets as long as they o not cause any ba effect on (non-etouring) shortest-path traffic. By protecting the shortest-path traffic from the harm of etouring traffic, we try to have the shortestpath traffic behave the same as in the non-loa-balance scheme, even in our loa balancing environment. Note that to reuce the ba effect (such as elay) on shortest-path traffic, it is important that T 2 is sufficiently small, while T 2 shoul hol minimum necessary value to cover the fluctuation of traffic istribution. C. Enabling Multiple Detour We further consier etouring packets more than once to improve the performance in larger networks. In a larger network, a packet meets congestion more than once with higher probability so that multiple etouring is essential from the viewpoint of scalability. To enable this, we introuce r-count instea of r-flag in packet heaer which inicates the number of etouring that the packet experience. By introucing the maximum allowe number of etouring for each packet, we allow several etouring while preventing packet loops. If we allow etouring three times, a 2-bit fiel is require for r-count. The change on router behavior is simple: if a packet without b-flag is receive an the primary next-hop is congeste (i.e., the queue length is more than T 1 ), the router forwars the packet into backup next-hop an increments the r-count of the packet only if which r-count is lower than the maximum allowe value, an the packet is roppe if the r-count is equal to the maximum allowe value. Note that only the packets which are using primary route (i.e., the packet without b- flag) can be etoure, because SBR oes not provie further backup paths of the packets which are alreay using backup paths. Table. I summarizes the whole behavior of routers in our proposal. This table shows the process of route ecision at each router when a packet arrive at the router. In etail, it presents the operation process on forwaring an flag manipulation for each state of packet heaer (b, r) an queue length. Here, let (b, r) be the b-flag an the r-counter of the packet, let max be the number of maximum allowe etouring, an let q (p) an q (b) be the queue length of the output queue for the primary an backup next-hop, respectively. III. TRAFFIC SIMULATION A. Simulation Setup We evaluate the propose loa balancing metho through traffic simulation using ns-2 [23] simulator with ranom topologies. To evaluate the performance of the case where traffic is partially concentrate on particular links, our scenario is esigne to locate several narrower links to cause congestion. Our simulation settings are shown as follows: We

4 Improvement of Throughput [%] 14.00% 12.00% % 6.00% 4.00% 2.00% Average Loa of Shortest-Path Packets per Link [%] Ratio of Detour Packets [%] 100% 90% 80% 70% 60% 50% num of etour: 3 num of etour: 2 40% num of etour: 1 30% 20% 10% 0% Increase of Packet Delay [%] 6.00% 5.00% 4.00% 3.00% 2.00% 1.00% num of noes: 30 num of noes: 100 Fig. 5. The Specification of Number of Detour to Fig. 3. Improvement of Throughput with Variation Reach Destination (Case of 30 Noes) Fig. 7. of Traffic Loa Traffic Increase of Packet Delay of Shortest-Path Ratio of Packet Rescue [%] Ratio of Detour Packets [%] 100% 90% 80% 70% 60% 50% num of etour: 3 num of etour: 2 40% num of etour: 1 30% 20% 10% 0% Ratio of Throughput [%] % % % 99.00% num of noes: 30 num of noes: 100 Fig. 4. Ratio of Packets Rescue by Detouring Fig. 6. The Specification of Number of Detour to Fig. 8. Reach Destination (Case of 100 Noes) Ratio of Throughput of Shortest-Path Traffic generate network topology of Waxman moel with 30 an 100 noes, respectively, using topology generator BRITE [20]. We select 10% of the links ranomly as bottleneck links which banwith is 50Mbps while normal link banwith is 100 Mbps. In this network, we generate several 10Mbps CBR flows with 1-kilobyte packets which source an estination noes are also selecte ranomly. We increase the CBR flows with time to raise the loa of the network. Output queue length is as long as 50 packets. Threshol T 1 is 100% an T 2 is 10% of the output queue length. The number of maximum allowe etouring is set to 3. Simulation is one by generating 10 topologies an for each topology we select ifferent two sets of bottleneck links an traffic patterns, an then we use average values to show the results. B. Simulation Results Simulation results are shown in Fig Fig. 3 shows the improvement of total network throughput of the propose metho compare to conventional single-path OSPF routing, with the variation of network loa. Network loa is inicate by average link usage of shortest-path packets, i.e., the packets which is not etoure. Note that since the number of flows transmitte is the same, the amount of shortest-path traffic is the same in both OSPF an the propose. The result shows that, in both 30 noes an 100 noes, the propose metho improve network throughput as the amount of traffic raise, an at 40-50% of link usage the improvement reach about 10%. There are a little ifference between the case of 30 noes an 100 noes, but totally they inicate the similar performance. Fig. 4 shows the ratio of packets rescue among all etoure packets. The both cases inicate the similar performance an when the link loa is 20-30% the rescue ratio is very high. Fig. 5 shows the specification of rescue packets in the number of etouring experience. The most part of them is rescue by one etouring while two an three times etoure packets occupy consierable part of them. Also, this result inicates that more than three etour will not be effective because it will rescue only a small number of packets. Fig. 6 shows the same specification in the case of 100 noes. The similar property is seen here except that the ratio of more than one etour goes larger. As for the ba effect of etouring traffic on shortest-path traffic, see Fig. 7 an Fig. 8. Fig. 7 shows that the elay of the shortest-path packets increase at most only 5%, which woul be the practical level making no problems. Note that we can ecrease the elay if we use smaller T 2 value, which is the trae-off between the elay an the buffer for the fluctuation of traffic. In Fig. 8 the ratio of shortest-path traffic throughput between OSPF an the propose metho are compare. In both cases of 30 noes an 100 noes, the throughput of shortestpath traffic is almost the same between OSPF an the propose metho, an this tren is the same regarless of the number of noes. From the whole results above, the propose metho rescues packets to improve at most 10% of throughput using vacant resources while they (the etour packets) o not make ba effect on the shortest-path traffic. The propose metho improves total throughput by 10% even when the loa is uniformly increase in the network.

5 IV. CONSIDERATION OF BASE IP FAST REROUTE MECHANISMS In this paper we investigate the performance of loa balancing metho which is base on an IP fast reroute mechanism SBR. There are, however, several well stuie fast reroute mechanisms [7]- [19], an esigning loa balancing mechanisms base on them is also possible. Here we escribe the ifference among them an iscuss the issues to moify them to provie loa balancing functions. In esigning IP fast reroute mechanisms, the traeoff between the overhea require to exten routing protocols an the capability of the scheme is important. Lee et al. propose failure inferencing base fast rerouting technique FIR [11]. FIR requires an aitional routing table for each network interface as overhea, while it is capable to recover every single link failure. Later, they propose FIFR [12] [13] which recover single noe failure with the same overhea. Shan et al. propose IP fast reroute techinique base on the approach calle NotVia aresses [9]. It recovers single noe/link failure by forwaring packets first to some intermeiate noes using IP tunnels, an then forwaring them to their estinations using shortest paths. In this scheme we nee overhea of IP encapsulation an the table of intermeiate noes for each estination. Kvalbein et al. propose a multi-tree approach of noe protection calle MRC [14] [15], which computes multiple spanning trees compute from the topologies in which several noes an links are isolate from the original topology. Those schemes recover every single noe/link failure an possibly multiple failures. As the overhea they require a mark on packet heaer to istinguish which tree the packet follows, an have to maintain multiple trees reay to forwar packets. To esign loa balancing mechanisms with those schemes, we have to implement 1) the counter of rerouting to prevent loops an 2) the flag to juge whether a packet is uner etour or not to give priority to the shortest-path traffic. Although small techniques may be require, basically it is not ifficult to implement them on the schemes above. As for the ifference between loa balancing performances, we say that the performance epens on backup paths compute by IP fast reroute schemes. Thus, although small variation will be seen, the link-protection scheme (e.g., FIR an NotVia with Link protection) base loa balancing mechanisms are likely to perform as well as the SBR base mechanism implemente in this work. Naturally, noe protection base loa balancing or multi-tree (MRC) base loa balancing will achieve better performance for hot spots covering wier area. V. CONCLUDING REMARKS In this paper we investigate a loa balancing technique base on IP fast reroute mechanisms. Our metho etours packets only when the packets meet congestion to make the most of vacant resources of networks. Through traffic simulation with increasing network loa, we foun that the propose metho improves network capacity at most 10% uner the scenario with congestion hot-spots. Also in low-loa state, significant ratio of packets is rescue by our scheme. As future work, it is essential to take TCP traffic into account. Specifically, since TCP performance is egrae ue to packet reorering, we have mainly two choices to eal with TCP traffic, i.e., 1) etouring TCP traffic such that packets in a flow use the same path, an 2) etouring only UDP packets in orer to avoi/issolve congestion hot spots. To fin the better choice is one of the important issues. REFERENCES [1] D.O. Awuche, J. Malcolm, J. Agogbua, M. O Dell, an J. McManus, Requirements for Traffic Engineering Over MPLS, IETF RFC2702, [2] B. Forts an M. Thorup, Internet Traffic Engineering by Optimizing OSPF Weights, in Proc. INFOCOM2000, pp , [3] S. Dasgupta, J.C.e Oliveira, an J.P. Vasseur, A Performance Stuy of IP an MPLS Traffic Engineering Techniques uner Traffic Variations, IEEE Globecom2007, pp , [4] A.K. Mishra an A. Sahoo, S-OSPF: A Traffic Engineering Solution for OSPF base Best Effort Networks, In Proc. IEEE Globecom 2007, pp , [5] M. Antić an A. Smiljanić, Oblivious Routing Scheme Using Loa Balancing Over Shortest Paths, In Proc. IEEE ICC 2008, pp , [6] E. Oki an A. Iwaki, F-TPR: Fine Two-Phase IP Rerouting Scheme over Shortest Paths for Hose Moel, IEEE Communications Letters, Vol.13, No.4, pp , [7] S. Rai an B. Mukherjee, IP Resilience within an Autonomous System: Current Approachs, Challenges, an Future Directions, IEEE Communications Magazine, Vol. 43, No. 10, pp , Oct, [8] S. Bryan an M. Shan, IP fast-reroute framework, IETF Internet-raft raft-ietf-rtgwg-ipfrr-framework-09.txt, [9] M. Shan, S. Bryan an S. Previi, IP Fast Reroute Using Not-via Aresses, raft-ietf-rtgwg-ipfrr-notvia-aresses-04.txt, [10] C. Reichert, Y. Glickmann an T. Mageanz, Two Routing Algorithms for Failure Protection in IP Networks, in Proc. of 10th IEEE Symposium on Computers an Communications (ISCC2005), pp , [11] S. Lee, Y. Yu, S. Nelakuiti, Z.L. Zhang, an C.N. Chuah, Proactive vs. Reactive Approaches to Failure Resilient Routing, in Proceings IEEE INFOCOM 04, Mar [12] Z. Zhong, S. Nelakuiti, Y. Yu, S. Lee, J. Wang, an C.N. Chuah, Failure inferencing base fast rerouting for hanling transient link an noe failures, in Proceeings of IEEE Global Internet, Mar [13] J. Wang an S. Nelakuiti, IP Fast Reroute with Failure Inferencing, In proceeings of SIGCOMM workshop (INM) 2007, pp , [14] A. Kvalbein, A.F. Hansen, T. Čičić, S. Gjessing, an O. Lysne, Multiple routing configurations for fast IP network recovery, IEEE/ACM Transactions on Networking, Vol.17, Issue. 2, pp , [15] T. Čičić, A.F. Hansen, A. Kvalbein, M. Hartmann, R. Martin an M. Menth, Relaxe multiple routing configurations for IP fast reroute, In proceeings of NOMS2008, pp , [16] T. Yoshihiro, A Single Backup-Table Rerouting Scheme for Fast Failure Protection in OSPF, IEICE Transactions on Communications, Vol. E91- B, No. 9, pp , [17] H. Ito, K. Iwama, Y. Okabe, T. Yoshihiro, Single backup table schemes for shortest-path routing, Theoretical Computer Science, 333(3): , [18] H. Ito, K. Iwama, Y. Okabe an T. Yoshihiro, Polynomial-Time Computable Backup Tables for Shortest-Path Routing, In Proceeings of SIROCCO 2003, , Jun, [19] K. Xi an J. Chao, IP Fast Rerouting for Single-Link/Noe Failure Recovery, In Proc. IEEE BROADNETS2007, pp , [20] A. Meina, A. Lakhina, I. Matta, an J. Byers, BRITE: An approach to universal topology generation, in Proceeings of IEEE MASCOTS, pp , Aug, [21] B. Waxman, Routing of Multipoint Connections, IEEE J. Select. Areas Commun., December [22] A.L. Barábasi an R.Albert, Emergence of Scaling in Ranom Networks, Science, 286: , October [23] The network Simulator NS-2,