SDN: learning notes for distributed and centralized control Layers

SDN: learning notes for distributed and centralized control Layers

"Water plays an important role in the development of human civilization. The history of computers used to store, process, and disseminate information and knowledge is similar to the history of storing, using, and managing water. In primitive tribes and nomadic periods, people's life and production are limited to the areas close to water sources. When water is needed, they need to fetch water from rivers and lakes. 'rivers and lakes 'are like 'hosts '. In the personal computer era, as in the farming age of mankind, because every household has water wells, people's life and production have gained more freedom and convenience. The 'waterwell 'is like a personal computer and self-sufficiency, however, it cannot be shared, a large number of idle resources, causing serious waste and inefficiency. When we arrived in the city, with a developed tap water network, the 'water' was concentrated for processing and provided to people anytime and anywhere. Therefore, civilization entered a new stage, just like "cloud computing" will lead us into a new era of Informatization with broadband networks and a new generation of software and services." This is a description of the development of the human information industry.

I think this section of the development history of computers is also a good description of centralized and distributed systems. It is a subtle idea to use water to compare computer resources. From the history of water resource processing, we can find an interesting phenomenon: People First rely on water sources, and then drill themselves to obtain water sources, at last, a unified factory was used to centrally manage water sources. This mode is like this for network storage: the earliest computers stored information in a unified manner. Users are influenced by network performance, and then Personal computers are developed, the entire network resources are distributed across various computer hard drives, and now the cloud-based storage is to centralize the entire resource, this is actually a process from centralized to distributed to centralized. It can be found that the concept of concentration and distribution often appears in resource management. When many resources are concentrated in one place, this is centralized. when resources are scattered in different places, this is distributed. They have their own advantages and disadvantages for centralized and distributed systems. It will be analyzed based on the actual network.

So how should we look at the centralized and distributed network for the entire network environment? Here, the centralized and distributed network refers to the structure of the control plane. It is a family chart of distributed options for control plane and data plane in SDN-Software Defined network.

From the spectrum above, we can see that the author believes that the network control plane system has three different models: strict centralization and semi-centralized or logical centralization, the last type is completely distributed.

ForStrictly centralized control planeHas the following features:

1. It is most commonly used in experimental SDN controllers;

2. Unified configuration platform;

3. single point of failure;

4. It is difficult to scale horizontally.

ForSemi-centralized or logic-centralized control planeAnd has the following features:

1. Typical solutions for modern SDN controllers;

2. Unified configuration plane, which needs to be synchronized with other control plane instances behind the scenes, but takes some time;

3. multiple points of failure can be recovered, but it is still vulnerable to the status synchronization with other control plane instances;

4. Easy to scale horizontally. You only need to deploy a new instance of the control plane.

ForCompletely Distributed Control PlaneAnd has the following features:

1. Typical methods;

2. Each (logical or real) device has a control plane instance;

3. It has been proved that the fault is highly recoverable;

4. There may be difficulties in convergence;

5. You need to configure and manage N instances;

6. It is difficult to scale out. New devices need to be added for scale-out.

In SDN, how should we determine the position of the control plane? Ken Gray proposed a very interesting concept: revolution, evolution, and tradition ).

So what is the revolutionary method? The revolutionary method proposed a new solution that completely adopted the centralized control plane of the network. In this model, the device does not have the control plane function, and the devices in this model are all silly and fast switching devices, completely controlled by the centralized control plane at the far end. Here, for the robustness problem, OpenFlow actually has a good solution, that is to say, multiple controllers are used at the same time (OpenFlow mentioned this scheme when the network element device establishes a communication with the Controller) to handle controller exceptions. For the coexistence of multiple controllers, OpenFlow provides two modes: Equal and Master/Slave ). The differences between the two modes will be explained in detail later on the OpenFlow protocol and SDN controller.

The evolution mode refers to the semi-centralized control plane model in the middle of the spectrum diagram, which focuses on the network domain defined in general. This mode can usually work with a distributed control plane in some form, which means that the device retains some traditional control plane functions (such as ARP processing or MAC address learning ), at the same time, a centralized controller is allowed to manipulate other functions, that is, those functions that are more convenient in the centralized operation paradigm. Such views are often treated as a hybrid mode or the concept of "bearing/overlapping, that is, the distributed control plane acts as the bearer, while the centralized control plane uses the bearer network for network transmission and logically overlays new functions. In this mode, we often need to consider the problem of how to implement "load/overlay "? Two different switches are defined in the OpenFlow white skin. One is the Dedicated OpenFlow Switch, and the other is the OpenFlow-Enabled Switch ). In fact, the compatible OpenFlow switch is the product under the evolution mode. The OpenFlow-compatible switch differs from the OpenFlow-dedicated switch in two aspects: first, the compatible switch has a Normal reserved Port (Resevered Port ); second, compatible vswitches support Packet Handling Methods for forwarding data packets from normal pipelines. Because this type of Switch only needs to be compatible with the OpenFlow protocol on a traditional switch, it can be used as an effective transition from completely distributed to strictly centralized.

The traditional method, as its name implies, is that dozens of control plane models currently used by most commercial switches are completely distributed. In this model, each device has at least one data plane and a complete control plane. In addition, each independent control plane in the model must work with other control planes to support a whole and runable network. Obviously, this solution is neither revolutionary nor evolutionary.

SDN has three concepts that most resonate with people: programmable, control plane and data plane separation, and centralized control model with instantaneous Network State management. Specifically, the establishment of the network control level system model is one of the core ideas of the entire network architecture after the control plane and the data plane are separated. The transformation from distributed to centralized means that the centralized control layer must solve its shortcomings and give full play to the advantages of centralized Chinese. For the centralized control layer, centralization at the control layer will inevitably Facilitate network managers to manage and configure the entire network, rationally mobilize network resources, further optimize the network, and improve the effective utilization of the network. At the same time, the advantages of resource centralization can be used to better implement network programmability. However, the disadvantage of its robustness needs to be properly protected to ensure the security at the control level, and multiple measures should be taken to solve its scalability problems.