Wherever I go and with whomever I talk in the academic networking community these days, the topic of hybrid packet/optical networks comes up. A very common question is: "What's this Hybrid Packet/Optical networking all about, and what does one look like?"

This article offers a very brief description of the motivation for hybrid networks and the motivation for studying and building them.

A hybrid packet/optical network for the purpose of this discussion is not simply a network with packet switches and optical paths in it. In fact all packet networks of any size today rely on optical elements for transmission..

To see the defining factor in a hybrid packet/optical network, think of the topology of today's packet networks: you have packet routers interconnected by optical paths. These packet routers are smart and make switching decisions, while the optical paths are provisioned over periods of months and years. While of course there are switching elements in the optical network, these switching elements are not visible to the end user because they are static. These switching units are visible to the operators of the network, but because they are static on any timescale of less than months, they are only significant from a network planning point of view. They are not significant from a traffic switching point of view or even from a traffic engineering point of view.

In a hybrid packet/optical network, the switching points in the optical transmission network, along with the packet routers, become topology-defining parts of the network. Where before the topology of the network was described as a set of packet routers connected by a set of transmission links, the topology is now described in terms of a set of packet routers attached to an optical subnetwork. Since the path topology of the optical subnetwork is dynamic, the topology of the packet network is also dynamic.

When we say "packet" in this case, we mean IP packets, and the Optical paths can actually mean wavelength paths or in some cases virtual carriers in a SDH/SONET network.

There are two major reasons for the current interest in hybrid packet/optical networks: the ability to create cut-through light-paths or circuits, and for core network optimization.

Cut-Through Paths

In some cases applications require an end-end path with very low jitter characteristics that may be difficult to guarantee across an arbitrary packet network. Alternatively, the end nodes may require a very large bandwidth guarantee for a fixed period of time, and rather than attempting to guarantee (for example) a 10Gbps path through the packet network it may be operationally simpler or just more economical to provision the path as a cut-through path on the optical subnetwork.

Core Network Optimisation

If you look at the topology of a large packet network and trace the path of a packet through the network, it is by no means uncommon for a packet to need to traverse tens of nodes from end to end. If, for example, there are 10 hops then the packet will traverse 20 router interfaces.

In normal networks, the core network capital expense is typically swamped by the cost of the access network. However, the cost of access interfaces per unit bandwidth is dropping faster than the cost of core network interfaces. As the speed and number of access interfaces increases, the cost of the core network equipment as a percentage of overall network cost is rising.

Keeping all other things constant, if it were possible to reduce the number of packet switched hops in the end-end path, the cost of the core network would be reduced, since the total number of packet interfaces in the network would be reduced. Of course this is not as straightforward as it sounds, because if it were then the optimal network would be a full mesh between the edge nodes, which it isn't. This is in fact an optimization problem that traffic engineers designing core networks do as a matter of course based on medium-long term average traffic statistics in the network.

If, however, the optical subnetwork was dynamic on a minute-by-minute scale, there would be the possibility of performing this optimization on a minute-by-minute basis rather than a month-by-month or annual basis. Doing traffic engineering optimization on shorter and shorter timescales will quickly be defeated by the law of diminishing returns in most core networks, but in cases (like many Research and Education Networks) where there are point-point flows in the multi-Gbps range (for example transfer of terabyte datasets from astronomical or particle physics instruments) that are added to the network for fixed periods of time, the requirement for packet interfaces can be reduced by allowing cut-through reservation.

Cut-through reservation can take two forms: In one case, the end user could reserve bandwidth at the edge node which is capable of delivering either packet paths or lightpaths. If the request is for such a large chunk of bandwidth that it becomes more economical to take a whole lightpath for a single end user, then a lightpath reservation is delivered. Alternatively, the edge node could offer packet services only, but when a reservation request or bundle of reservation requests representing a very large chunk of capacity to another edge node is received, it can request a cut-through path to the other edge node, effectively reducing the demand on the core.

In networks where this kind of "lumpy" reservation is prevalent, it is possible to reduce the dimensioning of the core network in line with the reduction in the expected peak demand on the core network. The size of that reduction is one of the interesting open questions in core network design.

Let's take a slightly more concrete look at hybrid packet/optical networks and how they relate to IP networks as they exist today.

Figure 1. Today's IP Network

Figure 1 above shows a typical IP network of today. At the edge of a network is an edge router which presents (and often defines) the service to the end user, who could be an individual, or could be an entire network.

Figure 2. First Generation Hybrid Network

The first generation hybrid packet/optical networks being built today are topologically not very different from any modern wide-area network. The big difference is that the underlying optical network is made visible to the end user, or at least transmission path services are made visible to the end user or to the edge router. It is now possible to "dial up" path services from the edge. The network, however, is still actually two separate networks offering the two different service types, with separate routing instances, separate control instances, and separate signaling systems. The end user needs to make the decision of which type of service is required and to request appropriately.

Figure 3. One Model for a Next Generation Hybrid network

In figure 3 above we see one possible model for the next generation hybrid packet/optical network. In this network we have some nodes that are hybrid nodes. One hastens to add that a hybrid node does not necessarily consist of a single device supporting both packet and path services. It is simply a node at which both kinds of service are switched, and could consist of multiple devices. The key difference is that for the combined topology there is a single routing instance covering all adjacencies, be they path or packet adjacencies.

Since a hybrid node has knowledge of both the packet and the path subnetwork, it is possible for a packet node (or a control system with the full topology) to serve requests for both types of service, or even to make a decision about which kind of service should be reserved for a particular request if it could be served by either service type.

There are of course many interesting open questions in the design of hybrid networks. The exact configuration and architecture of a hybrid node, how to decide when to use a cut-through path, how to maintain routing stability in the packet network when there is a closed-loop control algorithm potentially changing the packet topology in the face of shifting traffic patterns are just a few of the issues that designers and developers will have to face.