The Open Compute Project (OCP) was founded in 2011 with the goal of commoditizing servers. Nowhere in the OCP charter will you find the word “commoditize,” but the intent is obvious. A Facebook brainchild, OCP’s purpose was to lower the cost of server infrastructure while removing any of the proprietary technology that distinguished server providers from one another. In recent years, other “open” foundations have kicked off, including Open19 Foundation, a consortium founded by LinkedIn and Hewlett Packard Enterprise .
The creation of these different consortia is rooted in economics. Commoditization drives down cost. Lower cost means lower operational expenses (OpEx) for hyperscale datacenters like Facebook , LinkedIn , and other large datacenter operators in Silicon Valley, Wall Street, and beyond. While lower OpEx certainly equals higher margins, is this effort around commoditization actually good for the IT market? Is it good for you? Let’s explore this a little deeper.
What is a “commoditized” server?
When OCP started, one of the phrases often used to describe an open compute server was “vanity free design.” This means two things:
- No cosmetics: industrial designs are stripped down and very minimal. No fancy bezels, no fancy logos—just bare-bones server hardware that does its job.
- No proprietary management: all of the “secret sauce” that separates Hewlett Packard Enterprise, Dell EMC, Lenovo , and Cisco Systems is removed and replaced by open standards management controllers with open interfaces, ensuring seamless management across server vendors.
The notion of minimal cosmetics in a server design appeals to most IT professionals. We are typically the sort of people who value function over form, so stripping out cool bezel designs to save cost? I get it. Stripping out vendor specific management? Not so cool, but we’ll get to that in a little bit.
Original Design Manufacturers (ODMs) perform design services for server vendors. From motherboard layouts to mechanicals, ODMs work very closely with silicon providers (such as Intel and AMD ) and Original Equipment Manufacturers (OEMs) in designing product. Because of this deep domain knowledge, ODMs can quickly deliver a “generic” server to the market.
What makes a white box server good for hyperscale datacenters?
The real question is, “Why is hyperscale the right customer for white box servers?” Part of the answer (in case you missed this point in my opening) is cost. Another factor is the ability to operate without the strong support that a major server vendor provides. Hyperscale datacenters tend to rate very high in the capability maturity model, while smaller organizations rate lower.
Figure 1: IT Capability Maturity Model
Hardware matters less to a hyperscale datacenter. These datacenters tend to be highly automated and orchestrated. Servers are merely platforms for housing virtual machines, containers and applications—if that platform fails, those applications are moved to other resources automatically. Some days later, a datacenter technician will locate that rack among the thousands and replace the failed server.
Figure 2: Comparing traditional OEMs’ platforms with white boxes.
I’ve compared the two offerings based on what I would care about as an IT executive. Performance is largely dependent on the underlying CPU platform; however, I give a push to the OEMs given what they can do in systems design.
Aren’t the OEMs in cloud datacenters?
They are in some. Many OEMs have taken their industry standard rack servers and stripped them of manageability and other features to meet the requirements of the cloud providers. A good example of this is Hewlett Packard Enterprise with its Cloudline offering, or Dell EMC ’s Data Center Solutions (DCS) offerings. Both of these companies are part of OCP and other open compute consortiums.
The net-net: is compute really commoditized?
In short, it is not. Recent launches from AMD and Intel show that the CPU market is anything but commoditized. While both share the x86 instruction set, there are plenty of distinctions in microarchitecture, memory, security, and performance. Also, let’s not forget about IBM , Qualcomm , Cavium and other non-x86 architectures. The software-defined era of the datacenter is causing IT organizations to look at different alternatives in both the software and hardware ecosystem. This, in turn, will lead to datacenters that house workload specific server platforms to gain best performance, power, operational efficiency, and TCO.
While hyperscale datacenters may continue to promote open consortia to satisfy their own needs and OpEx gains, compute is far from commoditized. Those of us that manage somewhere south of 20,000 servers still benefit from the design, manageability, and support that comes from the likes of Hewlett Packard Enterprise , Lenovo , Dell EMC , and Cisco.