Striving for better performance and quick disaster recovery, organizations are shifting workloads between different geographical locations. Why not do the same for the sake of cost savings?

A group of researchers led by a Massachusetts Institute of Technology graduate student has published a study that suggests that companies operating IT infrastructures spread over large geographical areas can leverage energy price fluctuations in various areas to achieve significant cost savings.

Since actual energy cost varies from minute to minute and from place to place, depending on demand and a host of other factors, Internet-scale organizations can benefit from dynamically rerouting traffic to places where energy is cheaper at any given moment.

Given that a specific set of conditions is in place – the paper says – a company that implements the method can cut its electricity cost by at least two percent, which for large companies like Microsoft or Google can mean saving more than $1 million per year. According to the team’s rough “back-of-the-envelope” estimates, Google spends more than $38 million per year to power its servers and Microsoft spends about $2 million less than that.

To make its cost-saving estimates, the team used data it received from Akamai, a U.S. content delivery network. The researchers used 24 days’ worth of traffic statistics from Akamai servers in nine regions. They also analyzed hourly electricity prices from 29 locations collected over 39 months. The paper’s authors acknowledge that the savings they project are far from being exact and explain that their main purpose is to prove that electricity prices fluctuate enough to warrant exploring the idea further.

Necessary prerequisites
“The problem with this is there’s a number of things that need to be in place before this can work,” said Asfandyar Qureshi, the MIT graduate student who led the project. The three essential prerequisites are widespread adaptation of a time-of-day billing model by utilities, energy elasticity and implementation of a routing layer in the data center that reacts to price fluctuations.

The team defines energy elasticity as the difference between the energy servers consume when doing work and the energy they consume when idling. If a server uses about as much energy to stay idle as it does when it is working, shifting workloads from one location to another will not achieve much in terms of energy cost savings.

Transition to time-of-day billing will require major systemic changes to take place. The study is based on electricity prices on an open wholesale electricity market, while highly regulated U.S. utilities and their customers have fixed-price agreements. “For our approach to work this has to be renegotiated in some way,” Qureshi said.

Exposure to cost volatility makes companies nervous
Business customers are likely to be hesitant to undertake such renegotiations, since fixed prices add to predictability of their business models. “Right now, a lot of the data center operators like the fact that they have a fixed price and they’ll (ask) utilities: give us a fixed price and don’t expose us to price volatility. They don’t want to be at the mercy of the utility.”

Microsoft’s Principal Infrastructure Architect Christian Belady, while agreeing that “cost predictability is a good thing,” said that the long-term savings promised by the method may justify exposure to price volatility.

Organizations could gain further benefits from price fluctuations by making demand-response agreements with utilities, whereby they would quickly shift workloads at peak demand, reducing strain on the grid and getting compensated by utilities in return. The same technology layer deployed to respond to geographic and temporal price fluctuations could be leveraged for swift demand response.

Organizations could also make arrangements to sell electricity they saved by shifting workloads back into the grid, Qureshi said. “Being exposed to what’s generally considered a bad thing (price volatility) is a good thing if you can respond to it.”

Change in colocation billing model needed
An organization that does not operate its own data centers and houses its IT equipment in colocation facilities would face another hurdle if it wanted to take advantage of the approach the study suggests. Rates colocation customers pay do not change in response to changes in the amount of power they use.

In some states, such as California, it is illegal for colocation providers to charge their customers directly for the amount of power they use. To do that, they would have to register as utilities and be regulated as such, according to Mark Bramfitt, manager for energy efficiency incentive programs at PG&E, one of California’s largest utilities.

An assumption the study makes (and its authors acknowledge it) is that the market will not respond to the changes in demand patterns widespread adoption of the approach would cause.

“They make the assumption that if you do this energy-cost-chasing, it won’t have impact on the market price,” Belady said. “It’s essentially an arbitrage opportunity that they’re chasing and (after a while) the arbitrage opportunities disappear.”

Belady also pointed out that the calculations did not include additional compute capacity a system would need to shift workloads in real-time in response to price fluctuations.

Overall, Belady said the study was a “great first step” in stimulating industry interest and discourse about the opportunity the idea may present.

IBM Climate Stewardship and Energy Program Manager Jay Dietrich agreed. “The opportunity that they evaluated is in fact an opportunity but I think – to some extent – the difficulty is always in the detail,” he said.

One of the major hurdles to widespread adoption of the method Dietrich pointed out were data center operators’ availability requirements. Many operators would be hesitant to take on the risk of a failed workload shift, resulting in downtime. “The cost of an outage on a blown move will dwarf the electrical (cost) savings. If I was running a company, I might ask you to demonstrate it to me with some of my less critical applications.”

Related feature: Choosing data centre location
Related news: Dynamic IT load distribution can save millions in data center energy costs
Related analysis: Be realistic about data center energy saving options