Combinatorial design of multi-ring networks with combined routing and flow control

Abstract

In this paper we present a novel design technique for packet switched networks. The design is based on the construction of multiple virtual rings, which enjoy the one-bridge property: the path between any two nodes is either confined to a single ring or traverses exactly two rings (passing through a single bridge node). Our best designs are constructed by using finite generalized quadrangles of combinatorial design theory. We present novel routing and flow control protocols that capitalize on the one-bridge property of the multi-ring network. Our protocols ensure that (i) no loss due to congestion occurs inside a network, under arbitrary traffic patterns; (ii) all the packets reach their destinations within bounded time with low jitter; and (iii) the bandwidth is allocated fairly and no host is starved. We provide both a theoretical analysis and an extensive simulation-based performance evaluation of our protocols.

Introduction

The ever-increasing growth of data traffic and its bandwidth demand necessitates careful design and planning of the infrastructures of next generation networks. Traditional approaches to network design either face computationally hard optimization problems or use heuristic methods with approximate answers. The complexity increase in such approaches is partially due to dependency on traffic models and estimations. In contrast, combinatorial approaches to network designs [28], [30], [31] offer deterministic bounds on the maximum route length, and on the survivability, independent of the traffic characteristics. Ensuring traffic-independent properties is particularly important for data networks where traffic is bursty and unpredictable.

As the starting point of this paper, we introduce a new class of virtual-ring network designs. These networks enjoy the one-bridge property: the path between any two nodes is either confined to a single ring or traverses exactly two rings (passing through a single bridge node). Our best designs are constructed by using finite generalized quadrangles of combinatorial design theory.

Our goal is to show which routing, flow and access control protocols are best suited to such network topologies. We consider two basic approaches. With a unified approach, the protocols are tightly coupled with the network design, so they can take full advantage of its topological properties. With a decoupled approach, the protocols are standard WAN protocols, which are oblivious to the network design. Instead, they employ shortest path routing and window-based flow/congestion control. This paper tests the two alternatives and compares their performance.

We focus on the ring topology as a basic building block for several reasons. The ring topology is well studied, and is widely used in LAN and MAN environments. Its attractive symmetry and cyclic structure allow for simple, decentralized, fair, and congestion-free control protocols. And since a ring is a minimal biconnected graph it also provides survivability with minimal cost. For example, a multi-ring topology is the backbone topology of choice in the synchronous optical network standard [1]. Surveys on other uses of multi-ring networks are [8], [16].

Recently there has been a renewed interest in ring-based networks in the context of resilient packet rings (RPR) networks [5], [23]. RPR is a new data transport technology for metropolitan area networks (MAN), to be standardized as IEEE 802.17. An RPR is a ring-based architecture that consists of two counter-rotating rings. RPR generalizes the spatial bandwidth reuse by adapting mechanisms found in buffer insertion rings.

Networks with a simple linear topology, such as a bus or a ring [2], [10], [17] are not throughput scalable. In contrast, networks with an arbitrary topology may be throughput scalable but they are typically not congestion-free. The only exception we are aware of is the MetaNet architecture [19], [20] which provides ring properties on a LAN with an arbitrary topology, by embedding a virtual ring around a spanning tree of the network. The virtual ring emulates a full-duplex ring with spatial bandwidth reuse, like the MetaRing [6], [18]. Since the ring spans every node, the maximum length of routing is O(N) for an N-node network. Thus, scaling the MetaNet with a single virtual ring to a wide area network (WAN) may not be efficient.

Combinatorial design theory was first applied to multi-ring network designs in [30], [31]. The authors showed how to use balanced incomplete block designs (BIBDs) to obtain congestion-free networks with scalable throughput. The techniques resulted in networks in which the maximum route length and the maximum degree (number of rings a node belongs to) are both bounded by $\text{O}\text{(}\text{N}\text{)}$, where N is the number of nodes in the network. These bounds are similar to those of earlier approaches to multi-ring network design (e.g., chordal rings [22], [24] or ring-connected-ring [9]), with the additional property of congestion-free routing.

Our first contribution is the introduction of network designs with the one-bridge property. We prove a new trade-off between the node degree and the ring size for such networks. Then we introduce generalized quadrangles (GQ) and GQ-based network designs. Our GQ-based networks have a high level of path redundancy which translates to high survivability in the face of link and node failures, and allows easy load balancing among rings. And our networks require only O(N^4/3) links in total, which significantly improves the O(N^3/2) links of [30].

Our next contribution is a suite of new topology-aware protocols which we collectively call the VRing protocols. These protocols are tailored to virtual multi-ring networks which have the one-bridge property. Within each ring our protocols emulate a virtual slotted ring [3], [15]. Our routing and control protocols ensure that (i) the network is congestion-free, under arbitrary traffic patterns; (ii) all the packets reach their destinations within bounded time; (iii) the bandwidth is allocated fairly and no host is starved.

We evaluated the performance of our protocols by an extensive delay and throughput simulation study. We first tested the behavior of our topology-tailored VRing protocols on their own. Then, we compared the performance of VRing to an Internet-like suite of protocols which we call the INet protocols.

Our results indicate that the network access delay is substantially higher in the VRing than in the INet. However, once the packets enter the network, in the VRing they are routed with no loss, while in the INet they may be dropped and retransmitted multiple times due to congestion. Thus, our results show that the end-to-end network delay in the VRing is significantly lower than in the INet. Furthermore, we show that the variability of the network delay (and hence the jitter) in the VRing is quite small, and remains bounded even when the traffic intensity increases. In contrast, the network delay has large variability in the INet, and the variability grows rapidly with the traffic intensity. Finally, our results show that the network links are utilized more efficiently in the VRing.

We conclude that from the network provider’s perspective, the VRing provides much better network utilization, and allows the provider to give the network’s users QoS guarantees regarding available bandwidth, fairness, and jitter. Furthermore, for user applications that are more sensitive to jitter than to delay, such as audio and video transmissions, the VRing approach may be advantageous.

Organization: The rest of this paper is organized as follows. In Section 2 we describe the network design model, and introduce GQ designs. 3 The, 4 Flow control describe the VRing access control and flow control protocols. In Section 5 we analyze the VRing protocols’ properties. In Section 6 we present the results of a simulation-based performance evaluation of both the VRing and the INet. We conclude in Section 7.