The explosion of giant Web properties has server vendors building a new kind of machine that is stripped down to the bare essentials and optimized for cost- and energy-efficiency, analysts say.
The latest entry comes from HP, which on Wednesday introduced a line of x86 servers designed for what HP calls "extreme scale-out" environments. The HP ProLiant SL servers have a layout that lets fans run at lower speeds, and omit features that HP says often aren't required by large Internet companies, such as redundant power supplies and advanced management software.
But HP is not the pioneer in this scale-out server market, with companies such as IBM, Sun, Dell, Super Micro, and Rackable offering products of their own.
Each vendor has its own approach, but in general these are "systems optimized for large homogenous application scale-out deployments, an application that spans 1,000 servers," says Forrester Research analyst James Staten. "These are typically things delivered by Web services, or cloud services as the new buzz term goes."
One key factor is recognizing the server is no longer the point of reliability and availability, Staten says. A software layer -- such as a virtualization platform -- is needed to ensure the application survives the failure of any particular node, and "the server just needs to be as cheap and efficient as possible," he says.
HP claims its new products will let customers "cut acquisition costs by 10% and power draw by 28%, while doubling their compute density." These claims would be hard to verify, because HP has not revealed pricing of the servers.
But customers buying thousands of servers can reasonably expect up-front and operational costs to decline by tens of percentage points, says Illuminata analyst Jonathan Eunice. Besides improvements in power and cooling efficiency, new servers offered by HP and rivals "strip down traditional high-availability features and management features to optimize the cost and the supply chain for massive build-out," Eunice says.
This approach is similar to the thinking behind blade servers, which eliminate various components to save space and power costs. But these new scale-out servers may be most similar to what Google has built in its own server farms. Google executives recently explained in a blog post that "we strip down our servers to the bare essentials, so that we're not paying for components that we don't need. For example, we produce servers without video graphics chips that aren't needed in this environment."
Because hardware is never 100% reliable, "enterprises spend a lot of time and money on maintenance," the Google blog post continues. "In contrast, we expect the hardware to fail, and design for reliability in the software such that, when the hardware does fail, customers are just shifted to another server. This allows us to further lower the cost of our servers by using commodity parts and on-board storage. We also design the systems for easy repair such that, if a part fails, we can quickly bring the server back into service."
While Google builds its own servers, systems vendors believe there is a great opportunity in selling stripped-down servers to the rest of the Internet world, and perhaps to enterprises running large data centers.
IBM has the iDataPlex, which pushes two racks together, housing 84 servers and lowering costs by letting fans run at a lower speed and sharing power whips and cables. Scaling up simply requires purchasing another double-rack system, but Staten says IBM's approach is limiting in that it uses a unique rack and server form factor, preventing use of third-party hardware.Rackable, which is changing its name to SGI, is in the scale-out market with the ICE Cube containerized data center as well as the new CloudRack C2, a rack of servers designed for high density and energy efficiency.
Eunice says that Sun and Dell offer a few servers designed for large scale-out environments, but Staten points to Super Micro as being the furthest along technologically. Super Micro has built a "jigsaw motherboard" that can be reconfigured based on a specific customer's needs, Staten says.
If a customer doesn't need remote monitoring tools, or PCI slots, those components just get stripped away to reduce costs. "Based on specific requests, [Super Micro] can disassemble and reconfigure those jigsaw pieces into the kind of server motherboard a particular client wants, and then sell it to them in lots of 1,000," Staten says.
The new technologies illustrate the growing importance of the network, rather than the individual computer, Eunice says. With large Web properties buying tens of thousands of servers, they want a hardware design that reduces up-front capital costs and ongoing costs for management and power use. Cisco has recognized that the network and computer are moving closer together with its Unified Computing System, another product that can be seen as part of this trend, he says.
"More and more of IT is becoming networked IT," Eunice says. "More of the applications, more of the services, more of the presentation and access to IT happens over a network."
The IDG News Service contributed to this report.