An ever-increasing number of users and networked devices is creating an unprecedented volume of data being transported over global networks and processed in web-scale data centers. Switches and routers are at the heart of this infrastructure. Continuous innovation has always been the standard in routing and forwarding, from the early days of X.25, frame relay and token ring networks to present highly scalable IP/MPLS architectures. We can expect the speed of innovation to increase further as we move from legacy, monolithic devices to open networks utilizing SDN and NFV technology.
While the high-level direction seems clear, the details of network architectures and devices still need working out. And this is not an easy task as numerous stakeholders are involved in the discussion. Data center experts have quite different requirements than operational staff from communications service providers. The performance benefits of hardware-centric approaches needs to be balanced with the flexibility advantages of software-centric solutions.
The opportunity for innovation comes with some risk of confusion due to the different approaches to solving the same problem. Some examples of these approaches are compared and contrasted below.
Legacy switching and routing
Legacy switching and routing products combine a forwarding plan with a control plane, as shown in the figure below. External interfaces are standardized while the software and hardware architectures of the monolithic device are proprietary. Routing and signaling protocols of the control plane define the forwarding behavior, which is provided in the data plane by the forwarding engine.
Combining a Forwarding Plan
Legacy switching and routing products combine a forwarding plan with a control plane. (Source: Ulrich Kohn)
As protocols and interfaces are standardized, different suppliers' network elements should easily interoperate. In practice, the fast innovation with routing and forwarding protocols plus proprietary extensions results in poor efficiency of multivendor networks as the common, consistently implemented feature set of different products tends to be quite small. In addition, multivendor networks require extensive testing with each new release provided by one of the suppliers. Single-vendor technology domains became a common network design practice, creating a dependency of the service provider on a single supplier.
Monolithic, single-vendor solutions are a show-stopper on the transformation journey towards cloud-centric networking and network virtualization. Agile and flexible networks must rapidly move workloads to optimize resource efficiency, to support real-time provisioning of new services, and to meet aggressive new cost targets.
SDN promises improved scalability and better cost structure of connectivity networks by introducing network abstraction and separation of the control plane from the data plane.
The tight integration between forwarding and control plane is removed by an open interface using OpenFlow or NETCONF/YANG protocols. A central controller enables information dissemination and configuration of forwarding tables as it communicates and programs the simple switching devices through this open interface. There is less functionality in the switching device so the hardware becomes simpler. Open-source communities such as OpenDaylight or Open Floodlight are developing controllers, which also drives down cost.
The concept of abstractions enables decisions at a higher layer without a need to know all details of the network. Large-scale networks can be implemented without operating complex distributed signaling and routing protocols.
SDN has initially been developed for control of frame- and packet-forwarding hardware. Recently the applicability of SDN was extended toward circuit-switched technologies such as OTN or flexible photonic networks.
Instead of using a hardware-based forwarding engine, traffic switching and routing can also be done with router software running on standard processors. Virtual routers are in widespread use today in phones and personal computers for traffic forwarding between physical interfaces (Wi-Fi, Ethernet, USB) and software applications.
Virtualized switches and routers eliminate the need for dedicated hardware as those applications can run on standard servers, aka commercial off-the-shelf (COTS) servers. As those standard servers do not provide equivalent forwarding performance as dedicated hardware, various software technologies are developed for accelerating traffic handling. Examples include Data Path Development Kit (DPDK), single root - IO virtualization S(R-IOV), or vector packet processing (VPP).
If highest performance is required, the server CPU can be complemented by hardware functions and additional commands for offloading tasks such as encryption, identify management and data storage, among others. Such systems on chip (SoC) solutions close the gap between dedicated networking hardware and standard servers.
Open switches: Network operating system on bare metal
The next step in networking innovation is open switches running commercial or open-source network operating systems (NOS). A variety of suppliers build these open switches with similar features, sometimes based on open designs. This application is unlike SDN in that the control plane and forwarding plane are packaged in a common network element.
The open switch approach combines the low cost of mass production with the speed of software innovation from independent software solution providers. Open switches are deployed in data centers today. Those products are less prominent in public networks as they cannot provide similar sophistication with operational support features as legacy network elements.
How it all comes together
Switches and routers can be implemented in very different ways, which provides opportunities to optimize solutions for specific applications. Strengths and weaknesses of the different technical options must be mapped against application requirements. A short comparison of pros and cons is provided below, along with a helpful illustration.
Compare & Contrast
Service providers can implement switches and routers in very different ways, providing them with opportunities to optimize solutions for specific applications. (Source: Ulrich Kohn)
SDN nicely meets high-capacity connectivity demand in data centers. The lack of operational support capabilities does not make it a preferred technology in larger, geographically dispersed networks. There is increasing interest in using the technology in combination with other network control methods or as a control plane for circuit-switched high-bandwidth transmission technologies.
Open servers can host virtual switches and a wide range of other virtual network functions (VNFs) on open servers. This allows service providers to combine connectivity with other value-added services creating significantly more value than a proprietary switch or router. However, despite significant innovation with software acceleration technologies, virtual switches may not match the throughput and compute-intensive features of dedicated hardware appliances.
Bare-metal switches and NOS mainly address high port-count, high-capacity aggregation requirements of data centers. Service provider networks requires smaller size edge components as well, which is not in the focus of present bare-metal switches.
As any technology has some advantages and some disadvantages, I believe that a combination will result in the most efficient network devices optimized for specific tasks:
Data center switching will nicely combine bare-metal switches with thin NOS and central SDN technology for high throughput and holistic control
Edge switches with high performance requirements will be hybrid network elements combining open servers with open switches for best virtual switch performance and excellent hosting capability
Open server-based solutions satisfy low and medium performance requirements at the network edge. A single server can host a virtual switch in combination with further revenue generating software applications
SDN can be part of each of the above outlined solution. I predict SDN will become mainstream for control of circuit-based, high-bandwidth connectivity technologies such as OTN and photonic transport systems
Successful adoption of new technologies in the applications listed above will require collaborative work between various communities such as OPNFV, OpenSwitch, FD.io, ONOS/ONF and MEF among others, as well as the work of IETF and other specification-defining organizations. It will also require network operators and their suppliers to join forces in early technology evaluation, driving innovation from feasibility verification to network implementation, frequently transforming established processes to meet the agility, responsiveness and security requirements of cloud-centric networks.
This is not just the latest technology innovation -- we're re-inventing our networks.
— Ulrich Kohn, Director of Technical Marketing, ADVA Optical Networking. Special to New IP Agency.