Group leader election under link-state routing

Abstract

In this work, we place a long-established distributed computing problem in a new context. Specifically, the group leader election problem is studied “inside the network,” meaning that participants in the election process are network switches/routers, rather than hosts. In this context, an election protocol can take advantage of certain internal operations of the network, such as the underlying routing protocol, to meet stringent fault-tolerance criteria while minimizing the network traffic overhead.

A robust solution to the problem, called the Network Leader Election (NLE) protocol, is proposed. The protocol is designed for use in networks based on link-state routing (LSR). The protocol is robust, for it achieves leadership consensus in the presence of adverse events, such as leader failures and network partitioning. The correctness of the protocol is proved formally. A simulation study reveals that the NLE protocol incurs low overhead in handling leader failures and in-group creation. In addition, it is shown how important network functions, including hierarchical routing, address resolution, and multicast core management, can benefit from the NLE protocol.

Introduction

The problem of leader election concerns the selection of a distinguished member from a set of computing systems that are interconnected by a network. This problem has been extensively studied in the context of distributed computing systems, for example, in coordinating access to shared resources [1] and in implementing fault-tolerant objects [2]. Generally speaking, solutions to the problem are distributed “host-level” algorithms that make use of various network services, such as reliable delivery of messages, in order to monitor the working status of the established leader or cast ballots for a new leader. Well-known contributions in this area include the Bully algorithm [3] and the Ring algorithm [4]; more recent developments are described in Ref. [5].

In this paper, we address the leader election problem as it occurs “inside” the network. The participants in the election process are assumed to be switches (or, interchangeably in this work, routers), rather than hosts or application processes. Solutions to this problem are intended to support underlying network functions, as opposed to being directly invoked by user applications. Whereas a host-level election protocol typically considers the underlying network as a “black box,” we show that a network-level election protocol can access and take advantage of the internal operation of the network, in particular, the underlying routing protocol.

Network functions that can make use of an efficient leader election protocol are several. In this paper, we focus on three examples. First, in Asynchronous Transfer Mode (ATM) networks and other hierarchical networks, switches in a low-level subnetwork (called a routing domain) select a switch to represent the domain in the next routing level [6]; a solution to this domain leader election problem supports routing operations within the network. Second, many address-mapping services, such as the mapping between group addresses and member addresses [7] and the mapping between network addresses and link-layer addresses [8], use a central server approach; a solution to the server assignment problem selects a leader to undertake the server responsibilities. Third, some IP multicast protocols, such as CBT [9], identify a network node, called a core node, as the traffic transit center for each multicast group; a solution to this multicast core management problem supports multicast services provided by the network. A common requirement of solutions to the above problems is fault tolerance: since network functions/services are expected to survive not only single-point failures, but also component failures that partition the network, the solutions to these problems must also survive these adverse scenarios.

In this work, we explore the problem of leader election in networks based on link-state routing (LSR) [10], [11], an increasingly popular type of network routing. An LSR protocol makes complete knowledge of the network available to all switches. For this purpose, the local status of each switch, including the bandwidth available at incident links, buffer capacity, delays across links, and so forth, is learned by the network via the broadcasting, or flooding, of link-state advertisements (LSAs). Based on received advertisements, each switch locally maintains a complete topology image of the network, which it uses to make routing decisions. One of the major advantages of LSR is its fault tolerance. Since every link is monitored by its incident switches, and every switch is monitored by neighboring switches, malfunctioning components and congested areas are made known to all functioning switches promptly. Even the earliest LSR protocols were able to survive disastrous situations, such as network partitioning [11]. The Open Shortest Path First (OSPF) protocol [10], introduced by the Internet community, is one of the most well known LSR unicast protocols. LSR has also been adopted as the routing standard for ATM networks [6].

Our proposed solution, called the Network Leader Election (NLE) protocol, takes advantage of state information provided by an underlying LSR protocol. The NLE protocol also extends LSR to include group-leader binding LSAs, which are used by group members to advertise their choice of leader. Upon receiving such an LSA, other switches in the network either accept this selection, or choose and advertise an alternative leader. The objective of the NLE protocol is to achieve network-wide consensus on leader bindings. Specifically, we will show that the NLE protocol achieves the following properties:

• Leadership consensus property. Given a group G and a network that has been partitioned into a set of segments S₁,S₂,…,S_k, k≥1, there will be consensus on the leader binding for G within each segment S_i, and that leader will be an operational switch within the segment.

• Mutual consensus property. By requiring group members to report to the established leader, the NLE protocol ensures that, within each network segment S_i, the leader maintains a member list for the group that includes those, and only those, group members in S_i.

Simply put, the NLE protocol can handle leader failures and work properly under adverse scenarios such as network partitioning. When the network is not partitioned, the above consensus properties hold throughout the network. Results of a simulation study show that these features can be achieved with minimum protocol overhead.

The remainder of this paper is organized as follows. Section 2 reviews the operation of LSR and a previous network-level leader election method, the ATM domain leader election protocol which satisfies the two consensus properties discussed above. The design of the NLE protocol is presented in Section 3, and the correctness of the protocol, which is modeled as a consensus problem under LSR, is formally proved in Section 4. The performance of the NLE protocol and that of the ATM domain leader election protocol are compared via simulation in Section 5. Results of this study reveal that the NLE protocol incurs only a small fraction of the overhead of the ATM election protocol. In Section 6, we discuss the application of the NLE protocol to the address resolution problem and to the multicast core management problem; included are simulation results regarding the performance of NLE in creating multicast groups. Finally, conclusions are given in Section 7.