The new server refresh cycle: Why shorter really is better

By Jim O’Reilly, Contributor

Gains in performance, power, and density now follow Moore’s Law. Here’s how to take advantage of these new developments.

Until recently, an eight-year-plus life cycle has been the industry standard for computer equipment. The idea that hardware should last such a long time originated in the mainframe era and has continued to feed buying decisions, product designs, and data center planning.

However, the advent of inexpensive commercial off-the-shelf (COTS) hardware has changed the thinking on this process. Moore’s Law predicts that COTS CPU performance doubles roughly every 18 months, but that has had little impact on total system behavior, which has turned over on a much longer cycle. However, solid-state drives (SSD), new interfaces, and fast, large memories have all pushed system performance to get on board with Moore’s Law, too.

After an eight-year life cycle, COTS gear will be much slower and less efficient than new replacement servers. In fact, by Moore’s law, an 8-year-old system would be as much as 32 times slower than the latest servers on the market.

Another way of looking at this is to compare the cost of replacing servers more frequently with the cost of an eight-year replacement cycle. After eight years, based on this growth in horsepower, replacement could require only 3% to 6% of the number of existing servers. This change alone translates into major power savings.

Replacing hard drives with solid-state ones saves even more power. And if you adopt a cloud topology, you’ll get high utilization of the new servers, which may further reduce the number of servers needed. Such a large reduction in server count has implications for staffing as well.

Eight years is just too long to wait

Realistically, waiting eight years isn’t a good idea. Even a four to one improvement is sufficient to justify replacing after four or five years. This is because the replaced gear still will have residual value to offset the acquisition costs.

Replacing COTS gear is much easier than replacing legacy servers. Because the systems are essentially identical at the hardware abstraction layer, compatibility issues are extremely rare, and startup time is reduced. Additionally, modular or converged systems solutions let the COTS systems be purchased as pre-cabled, pre-integrated, and pre-tested clusters, taking weeks off startup time.

It’s worth noting that limiting a server’s working life to four or fewer years keeps it in the sweet spot of low annual failure rates. This is important for mission-critical systems, where downtime must be avoided at all costs.

The impact of early and efficient replacement is tremendous. As VP of engineering at Germane Systems, I helped build servers for the U.S. Navy’s Acoustic Rapid COTS Insertion (ARCI) program. The Navy contracted to deliver a new family of servers every two years, with two times the performance. Boats in the U.S. submarine fleet get updated every four years on a rolling schedule, boosting performance by at least four times with each update. One result of ARCI is that towed arrays now can extend to more than 32 miles. The same approach can be used in commercial organizations, where the idea of rolling out “pods” of gear every four years on a staggered schedule is making sense.

Looking at the bigger picture, many older configurations are built on servers with 1 GbE connections. In a virtualized data center with high performance storage and servers, today’s standard is 10GbE, which saves a large amount of money just by having 10GbE ports down on the motherboard.

Combine this approach with software-defined networking, which is taking the industry by storm, and a nimble high efficiency operation is within reach. Having seen the benefits of pairing virtual servers and software-defined networking, this, to me, is one of the strongest rationales for shortening server buy cycles.

About the Author: Power More