The Art of Network Architecture: Applying Modularity
- What Is Hierarchical Design?
- Typical Hierarchical Design Patterns
- Final Thoughts on Applying Modularity
Knowing and applying the principles of modular design are two different sorts of problems. But there are entire books just on practical modular design in large scale networks. What more can one single chapter add to the ink already spilled on this topic? The answer: a focus on why we use specific design patterns to implement modularity, rather than how to use modular design. Why should we use hierarchical design, specifically, to create a modular network design? Why should we use overlay networks to create virtualization, and what are the results of virtualization as a mechanism to provide modularity?
We’ll begin with hierarchical design, considering what it is (and what it is not), and why hierarchical design works the way it does. Then we’ll delve into some general rules for building effective hierarchical designs, and some typical hierarchical design patterns. In the second section of this chapter, we’ll consider what virtualization is, why we virtualize, and some common problems and results of virtualization.
What Is Hierarchical Design?
Hierarchical designs consist of three network layers: the core, the distribution, and the access, with narrowly defined purposes within each layer and along each layer edge.
Essentially, this definition takes one specific hierarchical design as the definition for all hierarchical design—we should never mistake one specific pattern for the whole design idea. What’s a better definition?
- A hub-and-spoke design pattern combined with an architecture methodology used to guide the placement and organizations of modular boundaries in a network.
There are two specific components to this definition we need to discuss—the idea of a hub and spoke design pattern and this concept of an architecture methodology. What do these two mean?
A Hub-and-Spoke Design Pattern
Figure 7-1 illustrates a hub-and-spoke design pattern.
Figure 7-1 Hub and Spoke Hierarchical Design Pattern
Why should hierarchical design follow a hub-and-spoke pattern at the module level? Why not a ring of modules, instead? Aren’t ring topologies well known and understood in the network design world? Layered hub-and-spoke topologies are more widely used because they provide much better convergence than ring topologies.
What about building a full mesh of modules? Although a full mesh design might work well for a network with a small set of modules, full mesh designs do not have stellar scaling characteristics, because they require an additional (and increasingly larger) set of ports and links for each module added to the network. Further, full mesh designs don’t lend themselves to efficient policy implementation; each link between every pair of modules must have policy configured and managed, a job that can become burdensome as the network grows.
A partial, rather than full, mesh of modules might resolve the simple link count scaling issues of a full mesh design, but this leaves the difficulty of policy management along a mishmash of connections in place.
There is a solid reason the tried-and-true hierarchical design has been the backbone of so many successful network designs over the years—it works well.
See Chapter 12, “Building the Second Floor,” for more information on the performance and convergence characteristics of various network topologies.
An Architectural Methodology
Hierarchical network design reaches beyond hub-and-spoke topologies at the module level and provides rules, or general methods of design, that provide for the best overall network design. This section discusses each of these methods or rules—but remember these are generally accepted rules, not hard and fast laws. Part of the art of architecture is knowing when to break the rules.
Assign Each Module One Function
The first general rule in hierarchical network design is to assign each module a single function. What is a “function,” in networking terms?
- User Connection: A form of traffic admission control, this is most often an edge function in the network. Here, traffic offered to the network by connected devices is checked for policy errors (is this user supposed to be sending traffic to that service?), marked for quality of service processing, managed in terms of flow rate, and otherwise prodded to ensure the traffic is handled properly throughout the network.
- Service Connection: Another form of traffic admission control, which is most often an edge function as well. Here the edge function can be double sided; however, not only must the network decide what traffic should be accepted from connected devices, but it must also decide what traffic should be forwarded toward the services. Stateful packet filters, policy implementations, and other security functions are common along service connection edges.
- Traffic Aggregation: Usually occurs at the edge of a module or a subtopology within a network module. Traffic aggregation is where smaller links are combined into bigger ones, such as the point where a higher-speed local area network meets a lower-speed (or more heavily used) wide area link. In a world full of high speed links, aggregation can be an important consideration almost any place in the network. Traffic can be shaped and processed based on the QoS markings given to packets at the network edge to provide effective aggregation services.
- Traffic Forwarding: Specifically between modules or over longer geographic distances, this is a function that’s important enough to split off into a separate module; generally this function is assigned to core modules, whether local, regional, or global.
- Control Plane Aggregation: This should happen only at module edges. Aggregating control plane information separates failure domains and provides an implementation point for control plane policy.
It might not, in reality, be possible to assign each module in the network one function—a single module might need to support both traffic aggregation at several points, and user or service connection along the module edge. Reducing the number of functions assigned to any particular module, however, will simplify the configuration of devices within the module as well as along the module’s edge.
How does assigning specific functionality to each module simplify network design? It’s all in the magic of the Rule of Unintended Consequences. If you mix aggregation of routing information with data plane filtering at the same place in the network, you must deal with not only the two separate policy structures, but also the interaction between the two different policy structures. As policies become more complex, the interaction between the policy spaces also ramps up in complexity.
At some point, for instance, changing a filtering policy at the control plane can interact with filtering policy in the data plane in unexpected ways—and unexpected results are not what you want to see when you’re trying to get a new service implemented during a short downtime interval, or when you’re trying to troubleshoot a broken service at two in the morning. Predictability is the key to solid network operation; predictability and highly interactive policies implemented in a large number of places throughout a network are mutually exclusive in the real world.
All Modules at a Given Level Should Share Common Functionality
The second general rule in the hierarchical method is to design the network modules so every module at a given layer—or a given distance from the network core—has a roughly parallel function. Figure 7-2 shows two networks, one of which does not follow this rule and one which does.
Figure 7-2 Poor and Corrected Hierarchical Layering Designs
Only a few connecting lines make the difference between the poorly designed hierarchical layout and the corrected one. The data center that was connected through a regional core has been connected directly to the global core, a user access network that was connected directly to the global core has been moved so it now connects through a regional core, and the external access module has been moved from the regional core to the global core.
The key point in Figure 7-2 is that the policies and aggregation points should be consistent across all the modules of the hierarchical network plan.
Why does this matter?
One of the objectives of hierarchical network design is to allow for consistent configuration throughout the network. In the case where the global core not only connects to regional cores, but also to user access modules, the devices in the global core along the edge to this single user access module must be configured in a different way from all the remaining devices. This is not only a network management problem, it’s also a network repair problem—at two in the morning, it’s difficult to remember why the configuration on any specific device might be different and what the impact might be if you change the configuration. In the same way, the single user access module that connects directly to the global core must be configured in different ways than the remaining user access modules. Policy and aggregation that would normally be configured in a regional core must be handled directly within the user edge module itself.
Moving the data center and external access services so that they connect directly into the global core rather than into a regional core helps to centralize these services, allowing all users better access with shorter path lengths. It makes sense to connect them to the global core because most service modules have fewer aggregation requirements than user access modules and stronger requirements to connect to other services within the network.
Frequently, simple changes of this type can have a huge impact on the operational overhead and performance of a network.
Build Solid Redundancy at the Intermodule Level
How much redundancy is there between Modules A and L in the network shown in Figure 7-3?
Figure 7-3 Determining Redundancy in a Partial Mesh Topology
It’s easy to count the number of links—but it’s difficult to know whether each path through this network can actually be considered a redundant path. Each path through the network must be examined individually, down to the policy level, to determine if every module along the path is configured and able to carry traffic between Modules A and L; determining the number of redundant paths becomes a matter of chasing through each available path and examining every policy to determine how it might impact traffic flow. Modifying a single policy in Module E may have the unintended side effect of removing the only redundant path between Modules A and L—and this little problem might not even be discovered until an early morning network outage.
Contrast this with a layered hub and spoke hierarchical layout with well-defined module functions. In that type of network, determining how much redundancy there is between any pair of points in the network is a simple matter of counting links combined with well-known policy sets. This greatly simplifies designing for resilience.
Another way in which a hierarchical design makes designing for resilience easier is by breaking the resilience problem into two pieces—the resilience within a module and the resilience between modules. These become two separate problems that are kept apart through clear lines of functional separation.
This leads to another general rule for hierarchical network design—build solid redundancy at the module interconnection points.
Hide Information at Module Edges
It’s quite common to see a purely switched network design broken into three layers—the core, the distribution, and the access—and the design called “hierarchical.” This concept of breaking a network into different pieces and simply calling those pieces different things, based on their function alone, removes one of the crucial pieces of hierarchical design theory: information hiding.
- If it doesn’t hide information, it’s not a layer.
Information hiding is crucial because it is only by hiding information about the state of one part of a network from devices in another part of the network that the designer can separate different failure domains. A single switched domain is a single failure domain, and hence it must be considered one single failure domain (or module) from the perspective of a hierarchical design.
A corollary to this is that the more information you can hide, the stronger the separation between failure domains is going to be, as changes in one area of the network will not “bleed over,” or impact other areas of the network. Aggregating or blocking topology information between two sections of the network (as in the case of breaking a spanning tree into pieces or link state topology aggregation at a flooding domain boundary) provides one degree of separation between two failure domains. Aggregating reachability information provides a second degree of separation.
The stronger the separation of failure domains through information hiding, the more stability the information hiding will bring to the network.