Authored by: Amit Katz – Sr. Director, Product Management
Customers are tired of paying huge sums of money for Ethernet switches for no good reason. At some point, OpenFlow seemed like the way to change the networking world, but various factors such as overlay networks, changing market interests, and other unforeseen developments, it is hard to view OpenFlow today as a game-changer. While it remains a very important technology and provides a valuable mean of implementing certain functionalities, it has not created a revolution in the networking industry.
The real revolution that is occurring today is based on a combination of the momentum gained by the Open Compute Platform and the increasing number of switch software and hardware suppliers. Initiatives to open the switch, such as Mellanox’s Open Ethernet that was announced earlier this year, have placed us on the right path to bringing networking to where servers are today: affordable, open, and software-defined.
But is this revolution all about saving on cost? Not at all – cost is important but flexibility, openness, and the freedom to choose are equally important. One of the key elements in enabling vendor selection elasticity is Open Network Install Environment (ONIE), which decouples the switch hardware from its software, enabling vendors to provide something very similar to what we see in the server world: hardware without an Operating System. That means customer can buy a server with many ports and install their choice of OS on top of it. In the event that the customer wants to change the OS, the lion’s share of the investment (the hardware piece) is protected.
Software-Defined Networking is all about being able to control the equipment you buy without needing to ask the vendor for new features. It is about controlling and deploying customer-specific features, improving existing mechanisms, and having the ability to actually program the network. This is similar to what is happening on the server side, where customers are writing their own applications. The best way to accomplish this is to use an open source switch, not to develop applications on top of a vendor’s proprietary/closed code.
So Open Ethernet is now a reality, but what exactly is Open Ethernet?
- The freedom to choose each of the following components:
- Switch silicon
- Operating System
- Software stack
- Ability to scale at lower costs (CapEx and OpEx)
Selecting the system itself is easy: the user chooses hardware based on cost, performance, scalability, reliability, redundancy, required certifications, and so on. Most of today’s discussions center around the readiness and robustness of the software/operating system, which is an essential building block toward Open Ethernet.
Mellanox recently announced the SX1024-OCP, an Open Compute Platform switch that provides a fresh alternative for deploying a non-blocking 10GbE rack with a reasonable price tag.
After selecting the various components, it is important to ensure that the promise of “scaling at the lowest cost” can be delivered, satisfying the rationale behind the networking revolution.
What makes a network more cost-effective?
- Switch silicon density is by far the most important CapEx factor.
- Switch silicon power consumption is the most important OpEx factor.
Once the software solution meets the user’s needs, switch silicon is probably the most important decision toward achieving the goal of scaling at lower cost
A fat tree architecture has become the standard way to build a high-performance, scalable data center. The following is a comparison between 36-port 40/56GbE and a 32-port 40GbE building block. Essentially, this compares Mellanox SwitchX-2 silicon vs. one of the alternatives.
* Mellanox 56GbE, when connected from leaf to spine enables larger fabrics, up to 756 external 40GbE ports.
Why is there such a vast difference? Why is the power consumption so disproportionate?
Mellanox has been building scalable fat tree, low power, low cost solutions for the past decade. Silicon density and efficiency was the only way to build very large, high-speed HPC fabrics, so the silicon is built for scale at cost.
Building a wire speed, high density 40GbE silicon is a challenge Mellanox successfully overcame two years ago, and today we understand the issue when we see alternative silicon that fails to reach line rate with 50% of the switching capacity. Failure to reach line rate with small packets can lead to significant network congestion, milliseconds of added latency, and poor application performance.
In summary, Open Ethernet has arrived, allowing customers to choose their switch silicon, hardware, operating system, and protocol stack for a reasonable price. Mellanox, together with other companies in the Ethernet industry, are leading this effort with SwitchX-2 and 40/56GbE silicon, thereby enhancing our customers’ ability to scale at linear cost and deliver the real promise of Software Defined Networking.