How open compute cuts server costs in the enterprise

The open compute project (OCP) means you can get the designs that Microsoft, Facebook and (to a lesser extent) Google use for their data centers.  The goal is to get original design manufacturers (ODMs) to build them for you rather than buying standard servers and switches from original equipment manufacturers (OEMs).

Six years on, it’s still mainly the largest of companies with the largest of data centers that are buying OCP designs, Forrester principal analyst Richard Fichera tells CIO.com. “Some of the larger financial services and large manufacturers are actively implementing OCP designs or OCP-like designs.”

Even so, he says, the project has had a big impact on the hardware that enterprises and mid-size companies buy. “Buying OCP-compliant systems is really still a high-end phenomenon but there’s been a more subtle ripple effect over the last few years that’s reached a long way.”

“When OCP started, almost all the major server vendors had what they called value product lines; systems that were stripped down and a little less expensive, with specs that were a little lighter all the way round, from connectors to power supply efficiency,” Fichera says.” OCP really highlighted a lot of the ways they could take further cost out of their servers and, by this point, all the major vendors have servers that have been influenced by the OCP designs. You can see that by the way they design systems and the resulting pricing of the entry level of value systems.”

But you won’t see those cost-cutting measures reflected in the list prices, Fichera says. “They don’t necessarily advertise as ‘this is a dirt cheap alternative to run your software’ but it gives them a tremendous amount of wiggle room when it comes to discounting.”

Bringing prices down was always part of the OCP plan, says Open Compute Foundation CEO Corey Bell. “We believe that everything should be more open because it allows for greater collaboration and communications; it drives down costs and it drives up effectiveness and efficiency. Where we used to be in the industry, vendors got input from customers but customers didn’t have a seat at the table to design products. Collaborating and designing openly is better for everyone, and it drives down costs.”

Enterprise scale, enterprise support

“A lot of the OCP designs were too wild and crazy for enterprise data centers, but they have some very mainstream 2U two socket designs with very standard server specs now and that’s trickled out into the industry,” says Fichera.

That’s exactly what Bell wanted to see happen. “We’re trying to make it easy for everyone to consume OCP and we’re pushing to overcome more of the hurdles. As we’ve brought on more component suppliers and more service providers as well as full products, we’re getting closer. When a product is submitted to OCP, we think about who would this work for and how would they get it.”

What OCP has already done is make cheaper systems available to everyone, albeit indirectly. “The difference between enterprise-class servers and a high-volume cloud server could be as much as 20 percent,” Fichera says, but he adds that “there’s now little difference in price between a ‘value’ server from the traditional server suppliers and an OCP design.”

What keeps many enterprises buying those value servers rather than OCP designs often has nothing to do with the hardware design, Fichera says. “Medium-size companies are still very sensitive to service arrangements,” he says, giving the example of companies buying several hundred servers who had been approached by server ODM Quanta with a slightly lower price than the traditional vendors, and turned down the deal.

“The rationale behind the decision is that they weren’t ready, for that price difference, to take the risk of what it means to be supported by a company like Quanta,” Fichera says. They fully understood that with a couple of hundred servers that they wouldn’t be getting the focused attention of Dell or HP [that a larger customer would get], but they were nervous about relying on a company with no visible presence in end user support.”

That’s an area OCP is working on, Bell tells CIO.com. “At the end of the day, it’s not just about deployment; for mid-range and smaller businesses, distribution and support is key. That’s where some of the major vendors have done a great job.” He notes that Dell and HP are already part of the OCP ecosystem, and Lenovo joined in early 2016. “We’re increasing and managing the ecosystem to provide that. There are going to be multiple players that enterprise will be able to go to,” he promises.

Cloud power

Even larger-scale systems like HPE’s Cloudline — which uses an OCP design rather than the proprietary technology used in ProLiant servers, and is manufactured in partnership with Foxconn rather than by HPE — are coming down market, says Fichera. “Initially they were very cagey about selling only to really big users but I get the impression they’re coming down now. If you want a couple of thousand units instead of ten thousand, I think they’ll talk to you now — and they have a lot of competition.”

While making data center power usage more efficient has been a big focus in OCP designs, that’s one area that isn’t directly relevant to enterprises (even though they typically pay more for power than cloud providers do). The biggest improvements require esoteric infrastructure like high-voltage DC power distribution that hyperscale cloud providers have adopted but few enterprises have in their data centers.

The good news is that both general customer demand and the lessons learned from OCP systems mean that server makers — and the power supply vendors they all use — have made their systems more energy efficient anyway. “Systems continue to exhibit big jumps in capacity in terms of performance per server; every generation of the base semiconductors has become more power efficient and both system vendors and CPU vendors have become more focused on power efficiency,” says Fichera.

In addition, many enterprises indirectly benefit from power efficiency improvements through cloud usage, Fichera points out. “Widely available cloud substitutes for a lot of workloads have reduced the number of people who wake up and have that ‘I have to build a new data center just to add more servers’ moment.”

Scaling up for private cloud

Even before OCP, the idea of buying barebones servers at scale appealed to some large businesses. Fichera knows of one large financial services company (he identified it only as one of the top 10 financial services organizations in the world) who approached an ODM in Taiwan. “They tucked one of their servers under their arm and said ‘Build me this but without this component, without that component, without this management agent; basically, build me a lower-cost version of this server.’ They bought probably 20,000 of them.”

Bell wants OCP to make that kind of custom manufacturing accessible to businesses of all sizes. “We want to make it easier to buy and use OCP, and to control your own destiny. With open collaboration, you can get exactly what you want, as opposed to being sold things you don’t need.”

That pick and mix approach is what Microsoft was doing with OEMs back in 2008 when it first announced Azure, and in 2011 it started doing much the same thing with Taiwanese ODMs like Quanta before it came up with its own server design, and joined OCP in 2014 to contribute that to the hardware community — and have it trickle down into the enterprise.

The most obvious way that trickling down is likely to happen is with Azure Stack, where Microsoft is specifying the hardware through the system vendors (OCP members Dell, HP and Lenovo). The complexity of servicing these systems is a big reason that pre-built systems will be the only option at first, Microsoft told us last year. The closer the hardware for Azure Stack is to the hardware used inside Azure, the easier servicing the systems is going to be — and Azure is built with OCP hardware.

But Microsoft’s latest OCP design might end up bringing you an OCP system more directly, suggests Kushagra Vaid, general manager of Azure hardware infrastructure at Microsoft. Project Olympus isn’t just an open specification; it’s also using the open design process familiar from open source. 

Last year, Microsoft published the Project Olympus spec, even though the design wouldn’t be finished until the middle of 2017. “We want to open up the design while it’s still in development so the community can come in and start to modify it and create derivatives. We can help drive standardization in the industry and reduce time to market for new offerings,” Vaid says.

Getting feedback and ideas before the design is set in stone will help make it applicable to a wider range of businesses, Vaid hopes. “Most OCP hardware works well for big companies but it doesn’t have enough diversity to be applicable for all scenarios and part of the reason for that is that when the design comes out, it’s too late to go make those changes. As the ecosystem gets built out, that should start to happen.”

The Project Olympus design covers everything from the motherboard, power distribution unit and server chassis interfaces to the rack itself, and they’re designed to be multi-purpose building blocks — and to let the OCP community add their own building blocks to the system.

Unusual for a cloud design, the motherboard isn’t specialized to a specific workload (which will also make it a better option for smaller scale). “It’s a universal motherboard designed to run any application you can think of. Today if you want to run a website, you pick one type of motherboard; if it’s for high-performance computing you pick a different motherboard, for machine learning it’s a different motherboard, for a database, it’s a different motherboard…. This is one motherboard that can address all those issues,” Vaid says.

You can plug in off-the shelf PCI Express cards for expansion, whether that’s a network card or an accelerator, and there’s an expansion board for adding GPUs, according to Vaid. “If you want a GPU, you add a GPU. You can add flash expansion, hardware expansion, video expansion…. That’s where you want to talk to the OCP community, to get those expan sion offerings.”

The power distribution unit is also universal, so it will work in different geographies with different power systems. “When you travel with your phone and you want to charge it, you plug it into an adapter,” Vaid says. “You don’t disassemble your phone every time you visit a new country.” Racks can be built centrally, shipped to where you need them — or moved from one data center to another, with the same power distribution strip, and you just plug in the adapter that’s right for the data center where you’re going to use it.

Vaid says he expects the design to be useful to private clouds like Azure Stack and data center system integrators as well as hyperscale clouds like Azure. “OCP solution providers can take these building blocks and customize them to make what data centers need. Today, what happens is that because they don’t have all the building blocks available in OCP, they go outside OCP and start building their own proprietary designs. They’re telling us that’s not part of the key value for them, but that they have no choice, and that if this ecosystem had the right pieces available already, they could focus on how you put it together.”

Bell echoes Vaid’s call for enterprises to get involved with the open design process, calling Project Olympus a great example of how OCP can improve for that audience. “Anybody can get involved in the community, and changes can be made by the big guys based on that. The more folks get engaged the more able we are to design hardware around their needs.

Bell also mentions Barreleye, Rackspace’s open server design for OpenPOWER (the technology alliance IBM founded in an attempt to break into the Intel-dominated data center market), noting that “Rackspace is an enterprise player, not a hyperscale cloud player and their spec fits the needs of that market.”

Making OCP relevant to businesses of all sizes improves the whole ecosystem, Bell says. “Everyone is looking to drive down cost of hardware. To design, produce and ship hardware is expensive, but now the cost of innovation is shared.”