A simulation study of the OSPF-OMP routing algorithm

Abstract

Open shortest path first (OSPF) is the most widely used internal gateway routing protocol on the Internet. However, one shortcoming is that it does not take advantage of the existence of multiple equal-cost paths between source and destination nodes. A well-known variation of OSPF, OSPF-ECMP (ECMP, equal-cost multipath), does exploit the presence of multiple equal-cost paths, but only on a static basis. A variation of OSPF, OSPF-OMP (OMP, optimized multipath), attempts to dynamically determine the optimal allocation of traffic among multiple equal-cost paths based on the exchange of special traffic-load control messages. This paper briefly describes the OSPF-OMP algorithm and the design of a discrete event simulator that models its behavior. We then use this simulator to carry out three experiments that compare the performance of OSPF, OSPF-ECMP, and OSPF-OMP under a range of traffic loads and distributions. Our results show that OSPF-OMP produces improvements in both delivery time and the number of lost messages when compared with the other two protocols.

Introduction

Open shortest path first (OSPF) is a link-state routing protocol developed by the Internet Engineering Task Force (IETF), and it is the internal gateway protocol currently recommended by the Internet Advisory Board [1], [2]. Like any link-state protocol, it may identify a number of distinct equal-cost paths between source/destination pairs. However, unless the protocol has been explicitly configured to take advantage of these multiple paths, it arbitrarily chooses one route and uses it in all forwarding operations.

One of the earliest attempts at exploiting equal-cost routes was the OSPF variant called OSPF-ECMP, an acronym for equal-cost multipath [3]. ECMP divides the total volume of traffic across all equal-cost paths using fixed, unchanging measures such as line speed or hop count. Referring to the sample four-node network in Fig. 1, if our distance metric is hop count, then there are two equal-cost paths from A to D––ABD with cost 2 and ACD, also with cost 2. OSPF would arbitrarily choose one of these two paths and use it for all traffic arriving at A destined for D. On the other hand, ECMP would split traffic equally between these two routes. This splitting process could be done using round-robin forwarding, in which messages 1,3,5,… go one way, while 2,4,6,… go the other, or via hashing in which the source and destination host identifiers are hashed to produce a numerical value that determines the specific route taken.

However, since ECMP does not dynamically adjust its division of traffic based on a knowledge of current loads, its allocation may be sub-optimal. For example, referring again to Fig. 1, assume that all links have capacity 1, the traffic volume from A to D is 1.2 units, from B to D, 0.5 units, and from C to D, 0.2 units. Using OSPF, all A to D traffic travels via a single route. If that chosen route is ABD, then the BD link is utilized 170% (1.2+0.5 units of traffic, with a link capacity of 1 unit) while the CD link is utilized only 20%. If OSPF instead chooses the ACD path, then the BD and CD links will be used 50% and 140% respectively. In both cases links are overloaded, and the system is unstable. With OSPF-ECMP, traffic from A to D is divided evenly between the two equal-cost routes. This produces utilizations along the BD and CD lines of 110% and 80% respectively––more closely balanced, but still unstable. An optimal split, based on given loads, would allocate 35% of the A to D traffic to path ABD and 65% to ACD, resulting in a utilization of 85% on both BD and CD––a stable system. Unfortunately, since ECMP does not dynamically examine loading values, it is unable to determine this optimal split.

There have been previous attempts to create adaptive routing algorithms that respond to changing traffic loads and patterns. Unfortunately, they have generally met with limited success due to oscillation and protocol overhead [4]. Referring once more to Fig. 1, if ABD were the optimal route from A to D, then a dynamic protocol would shift most traffic to it, causing increased delays and longer queues along that route. This will eventually cause ACD to become the better route, which will cause longer delays and queues along it, causing us to switch back again to ABD, etc. Dynamic algorithms are extremely sensitive to the “granularity” of change. That is, if the protocol discovers a better route but switches too much traffic too quickly, that route becomes overloaded and sub-optimal, quickly leading to oscillation and instability. On the other hand, if too little traffic is switched, then we are not exploiting the adaptive nature of the protocol and will not gain much when compared to traditional static algorithms.

A second major problem with dynamic routing methods is overhead. If we update too often then nodes will always have current loading information but at the cost of excessive network overhead. If updates are infrequent, then the loading information is out-of-date, and decisions about how to distribute traffic may be quite poor.

OSPF-OMP, an acronym for “optimized multipath” is the most recent attempt to create a load-sensitive routing algorithm. It was proposed by Curtis Villamizar of UUNET Corp. in October, 1997 and initially presented to the IETF in March 1998 [5]. Modifications to the original proposal were made in both 1998 and 1999, and the most recent IETF draft is entitled draft-ietf-ospf-omp-03, dated August 18, 1999, and available on the Web at http://www.brookfield.ans.net/ospf-omp/. For additional information about any aspect of OSPF-OMP, check the OMP home page at http://www.brookfield.ans.net/omp/.

This paper gives an overview of the OSPF-OMP protocol and then describes a discrete event simulation model created by the authors to analyze the performance of the entire OSPF family of routing algorithms. It then presents the results of experiments conducted with this model to study the performance of OSPF, OSPF-ECMP, and OSPF-OMP under both normal and highly stressed traffic loads. Our observations are compared with the results of earlier simulation experiments on OSPF conducted by the IETF. Our results demonstrate the important performance improvements that can potentially be achieved using the OSPF-OMP adaptive routing protocol.