Hop­by­Hop Routing Algorithms For Premium Traffic ∗ y Jun Wang Klara Nahrstedt Department of Computer Science Department of Computer Science University of Illinois at Urbana­Champaign University of Illinois at Urbana­Champaign Urbana, IL 61801, U.S.A. Urbana, IL 61801, U.S.A. [email protected] [email protected] ABSTRACT Differentiated Service (DiffServ) and hop-by-hop IP routing In Differentiated Service (DiffServ) networks, the routing assumptions. By combining service differentiation and QoS algorithms used by the premium class traffic, due to the routing together, the high-priority premium traffic should high priority afforded to that traffic, may have a significant be transmitted in an efficient manner with low negative in- impact not only on the premium class traffic itself, but on fluences to other low-priority traffic. Before presenting our all other classes of traffic as well. The shortest hop-count work in detail, some background knowledge is introduced in routing scheme, used in current Internet, turns out to be no the following subsections. longer sufficient in DiffServ networks. This paper studies the problem of finding optimal routes for the premium-class 1.1 Hop­by­Hop Shortest­path Routing traffic in a DiffServ domain, such that (1) no forwarding loop Hop-by-hop routing forms the basis of today's IP networks. exists in the entire network in the context of hop-by-hop Hop-by-hop routing means that routing decisions are made routing; and (2) the residual bandwidth on bottleneck links at each router independently and locally. For each incom- is maximized. This problem is called the Optimal Premium- ing packet at a router, its destination address (maybe some class Routing (OPR) problem. We prove in this paper that other fields in its IP header, too) is used to get the next hop the OPR problem is NP-hard. by consulting the router's routing table. Therefore, hop-by- hop routing is also referred to as destination-based table- To handle the OPR problem, first, we analyze the strength driven routing. The hop-count shortest-path routing (SP) and weaknesses of two existing algorithms (Widest-Shortest- is the most commonly used method for IP routing in today's Path algorithm and Bandwidth-inversion Shortest-Path al- Internet. Algorithms of finding the shortest-paths between gorithm). Second, we propose a novel heuristic algorithm, nodes, such as the Dijkstra's algorithm, can guarantee that called the Enhanced Bandwidth-inversion Shortest-Path (EBSP) no forwarding loop exists in a network. Figure 1 shows how algorithm. We prove theoretically the correctness of the the SP method works in a hop-by-hop scenario. EBSP algorithm, i.e., we show that it is consistent and loop- free. Our extensive simulations in different network environ- Routing Decision Routing Decision ments show clearly that the EBSP algorithm performs better Original Topology Made by A Made by B when routing the premium traffic in complex, heterogeneous A A A networks. 100 200 C B C Keywords B 10 C B Hop-by-hop Routing, Differentiated Service, Premium Class, Saturate Bandwidth. Routing Decision Resulting Topology with Packet Flows Made by C 1. INTRODUCTION A A Exploding traffic and expanding Quality-of-Service (QoS) 1 1 requirements, coming from emerging multimedia applica- 1 1 1 tions, initiated intensive research in QoS provision and rout- B C B C ing within the Internet. In this paper, we raise an interesting 1 problem of how to find optimal routing schemes under both Figure 1: An example of hop-by-hop SP routing: the ∗This work was supported by NSF Grant under contract number NSF ANI 00-73802 and NSF CISE Grant under numbers next to the links in the original topology contract number NSF EIA 99-72884. Any opinions, find- denote link capacities, while the labels in the result- ings, and conclusions or recommendations expressed in this ing topology mean how many “flows" go through material are those of the author(s) and do not necessarily the links. reflect the views of the National Science Foundation. y Please address all correspondences to Jun Wang and Klara The resulting topology, shown in Figure 1, illustrates the Nahrstedt at Department of Computer Science, University real packet flows1 between each pair of nodes. The num- of Illinois at Urbana-Champaign, Urbana, IL 61801, phone: (217) 244-5841, fax: (217) 244-6869. 1The word “flow" here is different from the definitions in ACM SIGCOMM Computer Communications Review 73 Volume 32, Number 5: November 2002 bers associated with the arcs are the numbers of flows going through that particular link in that direction. In this paper, we investigate a QoS routing problem in the 1 context of hop-by-hop routing, where bandwidth, rather Best−effort 0.8 than hop-count, is sensitive and the SP method turns out Assured to be insufficient. 0.6 Premium 0.4 0.2 1.2 DiffServ Model and Inter­class Effects Packet loss probability In order to provision better end-to-end QoS, DiffServ scheme 0 has been proposed as a cost-effective solution [6, 11]. In Diff- 0 1 Serv networks, traffic is classified into three service classes: 0.5 1 0.5 premium, assured and best-effort. The premium class traf- 1.5 fic has the highest priority in comparison to other classes Offered load (λ) 2 0 Fraction of premium traffic (p) of traffic. Originally, the DiffServ scheme is decoupled from IP routing intentionally, meaning that all traffic between (a) Packet Loss Probability each source-destination pair follows the same path no mat- ter which service class it belongs to and DiffServ itself has no effect on IP routing decisions. However, DiffServ does affect the queueing and drop behavior in routers. More specif- 100 ically, two separate queues are used in DiffServ. One of them is used by premium traffic and the other is shared by 80 non-premium traffic. The premium queue has absolutely Assured, Best−effort higher priority than the other queue, therefore, as long as 60 there are packets in the premium queue, these packets will be scheduled first. Due to premium traffic’s high priority, 40 Premium a bad routing decision could lead to some problems for the low-priority traffic when the volume of premium class traffic Packet delay (in packet) 20 is high. This means, without taking routing into considera- 0 tion, the premium class traffic imposes very negative influ- 0.2 1 0.4 ences on other classes of traffic, especially when the network 0.6 0.5 0.8 is highly loaded. We call this the inter-class effects [27]. In Offered load (λ) 1 0 Fraction of premium traffic (p) [21], the authors presented simple performance models and analysis of DiffServ schemes. However, to make a strong (b) Packet Delay (in packet) case of the negative impacts that the premium class traffic may impose on all other service classes in DiffServ networks, we have run simulations to measure the inter-class effects Figure 2: Measurements of the inter-class effects among all three service classes. Our results are shown in between all three classes in a DiffServ network Figure 2.2 As we can see in Figure 2, the premium class traffic has sig- nificant inter-class effects on the assured class and best-effort class traffic with respect to some important metrics, such as 1.3 Routing and Bandwidth Reservation for the packet loss probability (Figure 2(a)) and the packet de- Premium Traffic lay (Figure 2(b)). When the network is highly loaded (large During a certain time interval, presuming the full knowledge offered load λ) or the fraction of premium traffic is high of the network topology, we can regard the topology metrics (large p), the traffic with low-priorities experiences severe as static. Hence, we concentrate only on algorithmic aspects performance degradations (such as higher packet loss rates of static QoS routing schemes for the premium-class traffic and larger queueing delays). Therefore, we must take the in a DiffServ network. inter-class effects into consideration when we choose rout- ing algorithms for networks which support premium class In order to provide end-to-end QoS guarantees for the premium- traffic. class traffic between two nodes, certain amount of band- width must be successfully reserved on each link along the Integrated Service (IntServ) model or in RSVP. A “flow" path between these two nodes. (How to implement such here refers to a channel between two nodes, through which bandwidth reservation is out of the scope of this paper. We all premium traffic between the two nodes will follow. So assume that on top of our routing scheme, some class-based each pair of source and destination in the network identifies resource reservation protocol, similar to the RSVP, is used −! a “flow". For instance, Flow AB indicates the channel from for premium bandwidth reservation in the network.) If a A to B such that all premium packets from A to B are link is shared by multiple paths between different pairs of −! transmitted through AB. nodes, the bandwidth it has to reserve should accommodate 2More simulation details and extensive results and measure- the summation of the premium traffic on all the paths shar- ments are presented in our technical report [27]. ing this link. ACM SIGCOMM Computer Communications Review 74 Volume 32, Number 5: November 2002 D paths between nodes. More specifically, Bmax = maxfBs(R1); Bs(R2); · · · ; Bs(Rn)g; a (a+b) A B C where R1, R2, · · · , Rn are all possible loop-free routing b solutions for the given network. To find the optimal rout- a ing solution and the value of Bmax is called the Optimal Premium-class Routing (OPR) problem and it is not a trivial Figure 3: An example of bandwidth reservations for task.