SDN and Controller-Based Networks
Networking devices forward data in the form of messages, typically data-link frames like Ethernet frames. You have learned about how switches and routers do that forwarding for the entire length of preparing for the CCNA exam.
Network programmability and Software Defined Networking (SDN) take those ideas, analyze the pieces, find ways to improve them for today’s needs, and reassemble those ideas into a new way of making networks work. At the end of that rearrangement, the devices in the network still forward messages, but the how and why have changed.
This first major section explains the most central concepts of SDN and network programmability. It starts by breaking down some of the components of what exists in traditional networking devices. Then this section explains how some centralized controller software, called a controller, creates an architecture for easier programmatic control of a network.
The Data, Control, and Management Planes
Stop and think about what networking devices do. What does a router do? What does a switch do?
Many ideas should come to mind. For instance, routers and switches physically connect to each other with cables, and with wireless, to create networks. They forward messages: switches forward Ethernet frames, and routers forward packets. They use many different protocols to learn useful information such as routing protocols for learning network layer routes.
Everything that networking devices do can be categorized as being in a particular plane. This section takes those familiar facts about how networking devices work and describes the three planes most often used to describe how network programmability works: the data plane, the control plane, and the management plane.
The Data Plane
The term data plane refers to the tasks that a networking device does to forward a message. In other words, anything to do with receiving data, processing it, and forwarding that same data—whether you call the data a frame, a packet, or, more generically, a message—is part of the data plane.
For example, think about how routers forward IP packets, as shown in Figure 16-1. If you focus on the Layer 3 logic for a moment, the host sends the packet (step 1) to its default router, R1. R1 does some processing on the received packet, makes a forwarding (routing) decision, and forwards the packet (step 2). Routers R3 and R4 also receive, process, and forward the packet (steps 3 and 4).
FIGURE 16-1 Data Plane Processing on Routers: Basics
Now broaden your thinking for a moment and try to think of everything a router or switch might do when receiving, processing, and forwarding a message. Of course, the forwarding decision is part of the logic; in fact, the data plane is often called the forwarding plane. But think beyond matching the destination address to a table. For perspective, the following list details some of the more common actions that a networking device does that fit into the data plane:
De-encapsulating and re-encapsulating a packet in a data-link frame (routers, Layer 3 switches)
Adding or removing an 802.1Q trunking header (routers and switches)
Matching an Ethernet frame’s destination Media Access Control (MAC) address to the MAC address table (Layer 2 switches)
Matching an IP packet’s destination IP address to the IP routing table (routers, Layer 3 switches)
Encrypting the data and adding a new IP header (for virtual private network [VPN] processing)
Changing the source or destination IP address (for Network Address Translation [NAT] processing)
Discarding a message due to a filter (access control lists [ACLs], port security)
All the items in the list make up the data plane, because the data plane includes all actions done per message.
The Control Plane
Next, take a moment to ponder the kinds of information that the data plane needs to know beforehand so that it can work properly. For instance, routers need IP routes in a routing table before the data plane can forward packets. Layer 2 switches need entries in a MAC address table before they can forward Ethernet frames out the one best port to reach the destination. Switches must use Spanning Tree Protocol (STP) to limit which interfaces can be used for forwarding so that the data plane works well and does not loop frames forever.
From one perspective, the information supplied to the data plane controls what the data plane does. For instance, a router needs a route that matches a packet’s destination address for the router to know how to route (forward) the packet. When a router’s data plane tries to match the routing table and finds no matching route, the router discards the packet. And what controls the contents of the routing table? Various control plane processes.
The term control plane refers to any action that controls the data plane. Most of these actions have to do with creating the tables used by the data plane, tables like the IP routing table, an IP Address Resolution Protocol (ARP) table, a switch MAC address table, and so on. By adding to, removing, and changing entries to the tables used by the data plane, the control plane processes control what the data plane does. You already know about many control plane protocols—for instance, all the IP routing protocols.
Traditional networks use both a distributed data plane and a distributed control plane. In other words, each device has a data plane and a control plane, and the network distributes those functions into each individual device, as shown in the example in Figure 16-2.
FIGURE 16-2 Control and Data Planes of Routers—Conceptual
In the figure, Open Shortest Path First (OSPF), the control plane protocol, runs on each router (that is, it is distributed among all the routers). OSPF on each router then adds to, removes from, and changes the IP routing table on each router. Once populated with useful routes, the data plane’s IP routing table on each router can forward incoming packets, as shown from left to right across the bottom of the figure. The following list includes many of the more common control plane protocols:
Routing protocols (OSPF, Enhanced Interior Gateway Routing Protocol [EIGRP], Routing Information Protocol [RIP], Border Gateway Protocol [BGP])
IPv6 Neighbor Discovery Protocol (NDP)
Switch MAC learning
Without the protocols and activities of the control plane, the data plane of traditional networking devices would not function well. Routers would be mostly useless without routes learned by a routing protocol. Without learning MAC table entries, a switch could still forward unicasts by flooding them, but doing that for all frames would create much more load on the local-area network (LAN) compared to normal switch operations. So the data plane must rely on the control plane to provide useful information.
The Management Plane
The control plane performs overhead tasks that directly impact the behavior of the data plane. The management plane performs overhead work as well, but that work does not directly impact the data plane. Instead, the management plane includes protocols that allow network engineers to manage the devices.
Telnet and Secure Shell (SSH) are two of the most obvious management plane protocols. To emphasize the difference with control plane protocols, think about two routers: one configured to allow Telnet and SSH into the router and one that does not. Both could still be running a routing protocol and routing packets, whether or not they support Telnet and SSH.
Figure 16-3 extends the example shown in Figure 16-2 by now showing the management plane, with several management plane protocols.
FIGURE 16-3 Management Plane for Configuration of Control and Data Plane
Cisco Switch Data Plane Internals
To better understand SDN and network programmability, it helps to think about the internals of switches. This next topic does just that.
From the very first days of devices called LAN switches, switches had to use specialized hardware to forward frames, because of the large number of frames per second (fps) required. To get a sense for the volume of frames a switch must be able to forward, consider the minimum frame size of an Ethernet frame, the number of ports on a switch, and the speeds of the ports; even low-end switches need to be able to forward millions of frames per second. For example, if a switch manufacturer wanted to figure out how fast its data plane needed to be in a new access layer switch with 24 ports, it might work through this bit of math:
The switch has 24 ports.
Each port runs at 1 Gbps.
For this analysis, assume frames 125 bytes in length (to make the math easier, because each frame is 1000 bits long).
With a 1000-bit-long frame and a speed of 1,000,000,000 bits/second, a port can send 1,000,000 frames per second (fps).
Use full duplex on all ports, so the switch can expect to receive on all 24 ports at the same time.
Result: Each port would be receiving 1,000,000 fps, for 24 million fps total, so the switch data plane would need to be ready to process 24 million fps.
Although 24 million fps may seem like a lot, the goal here is not to put an absolute number on how fast the data plane of a switch needs to be for any given era of switching technology. Instead, from their first introduction into the marketplace in the mid-1990s, LAN switches needed a faster data plane than a generalized CPU could process in software. As a result, hardware switches have always had specialized hardware to perform data plane processing.
First, the switching logic occurs not in the CPU with software, but in an application-specific integrated circuit (ASIC). An ASIC is a chip built for specific purposes, such as for message processing in a networking device.
Second, the ASIC needs to perform table lookup in the MAC address table, so for fast table lookup, the switch uses a specialized type of memory to store the equivalent of the MAC address table: ternary content-addressable memory (TCAM). TCAM memory does not require the ASIC to execute loops through an algorithm to search the table. Instead, the ASIC can feed the fields to be matched, like a MAC address value, into the TCAM, and the TCAM returns the matching table entry, without a need to run a search algorithm.
Note that a switch still has a general-purpose CPU and RAM as well, as shown in Figure 16-4. IOS runs in the CPU and uses RAM. Most of the control and management plane functions run in IOS. The data plane function (and the control plane function of MAC learning) happens in the ASIC.
FIGURE 16-4 Key Internal Processing Points in a Typical Switch
Note that some routers also use hardware for data plane functions, for the same kinds of reasons that switches use hardware. (For instance, check out the Cisco Quantum Flow Processor for interesting reading about hardware data plane forwarding in Cisco routers.) The ideas of a hardware data plane in routers are similar to those in switches: use a purpose-built ASIC for the forwarding logic, and TCAM to store the required tables for fast table lookup.
Controllers and Software-Defined Architecture
New approaches to networking emerged in the 2010s, approaches that change where some of the control plane functions occur. Many of those approaches move parts of the control plane work into software that runs as a centralized application called a controller. This next topic looks at controller concepts, and the interfaces to the devices that sit below the controller and to any programs that use the controller.
Controllers and Centralized Control
Most traditional control plane processes use a distributed architecture. For example, each router runs its own OSPF routing protocol process. To do their work, those distributed control plane processes use messages to communicate with each other, like OSPF protocol messages between routers. As a result, traditional networks are said to use a distributed control plane.
The people who created today’s control plane concepts, like STP, OSPF, EIGRP, and so on, could have chosen to use a centralized control plane. That is, they could have put the logic in one place, running on one device, or on a server. Then the centralized software could have used protocol messages to learn information from the devices, but with all the processing of the information at a centralized location. But they instead chose a distributed architecture.
There are pros and cons to using distributed and centralized architectures to do any function in a network. Many control plane functions have a long history of working well with a distributed architecture. However, a centralized application can be easier to write than a distributed application, because the centralized application has all the data gathered into one place. And this emerging world of software-defined architectures often uses a centralized architecture, with a centralized control plane, with its foundations in a service called a controller.
A controller, or SDN controller, centralizes the control of the networking devices. The degree of control, and the type of control, varies widely. For instance, the controller can perform all control plane functions, replacing the devices’ distributed control plane. Alternately, the controller can simply be aware of the ongoing work of the distributed data, control, and management planes on the devices, without changing how those operate. And the list goes on, with many variations.
To better understand the idea of a controller, consider one specific case as shown in Figure 16-5, in which one SDN controller centralizes all important control plane functions. First, the controller sits anywhere in the network that has IP reachability to the devices in the network. Each of the network devices still has a data plane; however, note that none of the devices has a control plane. In the variation of SDN as shown in Figure 16-5, the controller directly programs the data plane entries into each device’s tables. The networking devices do not populate their forwarding tables with traditional distributed control plane processes.
Figure 16-5 shows one model for network programmability and SDN, but not all. The figure does give us a great backdrop to discuss a few more important basic concepts; in particular, the idea of a southbound interface (SBI) and northbound interface (NBI).
FIGURE 16-5 Centralized Control Plane and a Distributed Data Plane
The Southbound Interface
In a controller-based network architecture, the controller needs to communicate to the networking devices. In most network drawings and architecture drawings, those network devices typically sit below the controller, as shown in Figure 16-5. There is an interface between the controller and those devices, and given its location at the bottom part of drawings, the interface came to be known as the southbound interface, or SBI, as labeled in Figure 16-5.
Several different options exist for the SBI. The overall goal is network programmability, so the interface moves away from being only a protocol. An SBI often includes a protocol, so that the controller and devices can communicate, but it often includes an application programming interface (API). An API is a method for one application (program) to exchange data with another application. Rearranging the words to describe the idea, an API is an interface to an application program. Programs process data, so an API lets two programs exchange data. While a protocol exists as a document, often from a standards body, an API often exists as usable code—functions, variables, and data structures—that can be used by one program to communicate and copy structured data between the programs across a network.
So, back to the term SBI: it is an interface between a program (the controller) and a program (on the networking device) that lets the two programs communicate, with one goal being to allow the controller to program the data plane forwarding tables of the networking device.
Unsurprisingly, in a network architecture meant to enable network programmability, the capabilities of the SBIs and their APIs tell us a lot about what that particular architecture can and cannot do. For instance, some controllers might support one or a few SBIs, for a specific purpose, while others might support many more SBIs, allowing a choice of SBIs to use. The comparisons of SBIs go far beyond this chapter, but it does help to think about a few; the second major section gives three sample architectures that happen to show three separate SBIs, specifically:
OpenFlow (from the ONF; www.opennetworking.org)
OpFlex (from Cisco; used with ACI)
CLI (Telnet/SSH) and SNMP (used with Cisco APIC-EM)
CLI (Telnet/SSH) and SNMP, and NETCONF (used with Cisco Software-Defined Access)
The Northbound Interface
Think about the programming required at the controller related to the example in Figure 16-5. The figure focuses on the fact that the controller can add entries to the networking device’s forwarding tables; however, how does the controller know what to add? How does it choose? What kind of information would your program need to gather before it could attempt to add something like MAC table entries or IP routes to a network? You might think of these:
A list of all the devices in the network
The capabilities of each devices
The interfaces/ports on each device
The current state of each port
The topology—which devices connect to which, over which interfaces
Device configuration—IP addresses, VLANs, and so on as configured on the devices
A controller does much of the work needed for the control plane in a centralized control model. It gathers all sorts of useful information about the network, like the items in the previous list. The controller itself can create a centralized repository of all this useful information about the network.
A controller’s northbound interface (NBI) opens the controller so its data and functions can be used by other programs, enabling network programmability, with much quicker development. Programs can pull information from the controller, using the controller’s APIs. The NBIs also enable programs to use the controller’s capabilities to program flows into the devices using the controller’s SBIs.
To see where the NBI resides, first think about the controller itself. The controller is software, running on some server, which can be a VM or a physical server. An application can run on the same server as the controller and use an NBI, which is an API, so that two programs can communicate.
Figure 16-6 shows just such an example. The big box in the figure represents the system where the controller software resides. This particular controller happens to be written in Java and has a Java-based native API. Anyone—the same vendor as the controller vendor, another company, or even you—can write an app that runs on this same operating system that uses the controller’s Java API. By using that API to exchange data with the controller, the application can learn information about the network. The application can also program flows in the network—that is, ask the controller to add the specific match/action logic (flows) into the forwarding tables of the networking devices.
FIGURE 16-6 Java API: Java Applications Communicates with Controller
Before leaving the topic of NBIs, let me close with a brief explanation of a REST API as used for a controller. REST (Representational State Transfer) describes a type of API that allows applications to sit on different hosts, using HTTP messages to transfer data over the API. When you see SDN figures like Figure 16-6, with the application running on the same system as the controller, the API does not need to send messages over a network because both programs run on the same system. But when the application runs on a different system somewhere else in the network other than running on the controller, the API needs a way to send the data back and forth over an IP network, and RESTful APIs meet that need.
Figure 16-7 shows the big ideas with a REST API. The application runs on a host at the top of the figure. In this case, at step 1, it sends an HTTP GET request to a particular URI. The HTTP GET is like any other HTTP GET, even like those used to retrieve web pages. However, the URI is not for a web page, but rather identifies an object on the controller, typically a data structure that the application needs to learn and then process. For example, the URI might identify an object that is the list of physical interfaces on a specific device along with the status of each.
FIGURE 16-7 Process Example of a GET Using a REST API
Software Defined Architecture Summary
SDN and network programmability introduce a new way to build networks. The networking devices still exist and still forward data, but the control plane functions and locations can change dramatically. The centralized controller acts as the focal point, so that at least some of the control plane functions move from a distributed model to a centralized model.
However, the world of network programmability and SDN includes a wide array of options and solutions. Some options pull most control plane functions into the controller, whereas others pull only some of those functions into the controller. The next section takes a look at three different options, each of which takes a different approach to network programmability and the degree of centralized control.