Like far too many things in this world, enterprise networking seems to bounce between two extremes. One year, hardware acceleration is all the rage. The next, a software-only approach seeks to transform the way networks are built and deployed altogether. In reality, new networking approaches such as software-defined networking (SDN) and networking functions virtualization (NFV) will require a balance between hardware and software.
Three forms of silicon have dominated the network landscape. First, there are proprietary ASICs and processors that have historically been key to network hardware. More recently, commodity network processors from the likes of Broadcom and Intel have played a larger role. In general, the Broadcom processors tend to be more optimized for processing network services. But Intel, in the form of startup vendors such as 128 Technology and Netronome, is also making a compelling case for leveraging multicore processors to run network services next to compute and storage engines running on the same x86 processor.
Hardware Acceleration & Programmability
In recent weeks, a variety of startups have emerged that aim to strike a balance between hardware and software to build a modern network. Barefoot Networks unveiled a switch, slated for next year, that promises to deliver 6.5 Tb/s throughput by primarily relying on a custom Tofino processor that exploits the P4 programming language to create a SDN environment. The end goal is to not only scale networking performance to new heights, but also reaffirm the role proprietary processors can play in enabling those capabilities. This is a new trend toward a hybrid process – one that combines the performance of high-end proprietary chips with the programmability of network processors designed for an SDN world.
“The problem with most of the switches deployed today is that they are fixed rather than programmable,” says Ed Doe, vice president of products and marketing for Barefoot. “P4 allows us to separate the network data plane.”
At the other end of the processor spectrum is Big Switch Networks, which recently announced it is taking advantage of the Broadcom Trident-II+ and Tomahawk processors to double the number of leaf switches its can support using its SDN environment, to 64, inside a data center pod.
Based on a leaf-spine architecture, Big Switch believes the update to the company’s Big Cloud Fabric illustrates how quickly commodity processors from Broadcom are evolving.
“We can now support up to 24,000 virtual machines,” says Gregg Holzrichter, chief marketing officer for Big Switch Networks. “Before long we think you’ll be seeing NFVs everywhere.”
Hardware Abundance Spawns Network Startups
Of course, there are those who contend that advances in commodity hardware should also dictate further innovations at the software level that transform networking altogether. For example, 128 Technology created a session-oriented router based on Intel processors. CEO Andy Ory says 128 Technology is taking advantage of Intel processors to create a router that eliminates much of the software and appliances such as firewalls that are deployed to augment routers today.
The product is dubbed the 128T Networking Platform. Ory says routing and, by extension, the Internet itself, are broken because of all the complexity traditional routers introduce.
“We take routing all the way to virtual machines running on the core,” says Ory. “We’re going to eliminate all the needs for tunnels and network overlays.”
At the same time, Ory says the 128T Networking Platform will eliminate the need for MPLS networks, because IT organizations will have more granular control over the quality-of-service (QoS) associated with any session.
Of course, 128 Networking isn’t the only networking vendor looking to take advantage of x86 processors. Netronome, for example, has a designed a co-processor based on Intel processors that is housed on a card that plugs directly into a server. The basic argument is the card provides the dual benefit of offloading networking from the server in ways that still make it possible to exploit the economics of Moore’s Law.
Hardware Acceleration Debate Accelerates
The variety of new ways to build a network that can balance performance with programmability is likely to spur debate over which approach is best — whether it’s pure commodity hardware or selective hardware acceleration. The blending of purpose-built silicon with software-defined programmability, including the use of the emerging P4 language, appears to be a new trend. Other startups such as Netronome are looking at ways of combining programmability and targeted hardware acceleration to boost performance.
No matter the approach to networking, innovation advances in hardware and software need to appear hand-in-glove before most IT organizations can really tap their power. For the last few years, the networking innovation pendulum has been swinging towards software. But it’s also clear that it may now be starting to swing back toward both dedicated and commercial processors. Somewhere in the middle of those two extremes, there is an opportunity for IT organizations to transform today’s fairly rigid networking environments into something that can scale in ways that do more to increase overall IT agility than to compromise it.