Really Remote Data
Far-flung data centers could use otherwise unharvestable renewable energy for computation.
Researchers at Cambridge University want to put data centers in places so remote they aren’t on any power grid. Their models indicate that moving data-hungry computation to places such as scorching deserts, windswept peaks, and the middle of the Atlantic Ocean—all rich in sunlight and wind energy—could allow this otherwise unharvestable energy to do useful work.
In a paper to be delivered at the 13th annual HotOS conference in May, the authors offer an extreme model of how cloud services could incorporate remote data centers powered only by renewable energy. Their scenario sites one solar- and wind-powered data center in the desert of southwest Australia and a second one in Egypt, on other side of the planet. This placement is no accident: putting them in different hemispheres, on opposite sides of the earth, maximizes the solar and wind energy they can harvest.
One catalyst for such a radical rethinking of how data centers can be sited and powered is the increasing availability of advanced fiber-optic networks. Connecting a remote renewable-energy plant to a power grid remains prohibitively expensive, reasoned the researchers working on this project—Sherif Akoush, Ripduman Sohan, Andrew Rice, Andrew W. Moore, and Andy Hopper—but running fiber-optic cable to such a plant would be relatively easy and cheap.
“We envisage data centers being put in places where renewable energy is being produced and you could never economically bring it back to heat a house,” says Andy Hopper, senior author on the paper and head of Cambridge University’s computer science department. “But you could lay a fiber and use energy that is otherwise lost, in that it’s not economically transportable.” One way to think of the underlying principle, he notes, is that it’s easier to move bits (made up of photons) than electrons.
Jonathan Koomey, a researcher and consulting professor at Stanford, cautions that a number of real-world factors could render the Cambridge team’s hypotheticals invalid. While data centers are costly, Koomey explains, the value they create is so far in excess of those costs that anything that reduces their effectiveness would reduce their net benefit to society.
“If the actions you take to save costs would also cut into the number of computations that you can then deliver, you’ll reduce economic benefits from data centers, and that’s presumably not what the authors had in mind,” says Koomey.
Hopper, however, points out that the larger effort of which this paper is a part—the Computing for the Future of the Planet project—takes it as a given that more computing is always good, because the virtualization of goods and services displaces more energy-intensive activities in the physical world. He says that a system like the one he proposes would be implemented only at either “no cost to overall performance [of a cloud computing system] or at an attractive cost to performance.”
The key to incorporating far-flung, intermittently available data centers into a cloud infrastructure, says Ripduman Sohan, a postdoctoral fellow who worked on the paper, is to be choosy about which processes are offloaded to them.
“I think Facebook would not want to put forward-facing Web services onto an architecture like this, for various reasons,” says Sohan. However, a company like Facebook could offload projects that are not particularly time-sensitive but still sizable, such as processing analytics. “There are a bunch of batch-type jobs that could easily be offloaded to an architecture like this,” says Sohan.
At least one real-world implementation of a system similar to the one proposed by the Cambridge team already exists. Called the GreenStar Network, it connects data centers powered entirely by renewable energy in Canada, Spain, Ireland, and Iceland. So far, the challenges inherent in porting large amounts of data and live computing processes from one data center to another in near-real time have been significant but surmountable.
The network uses supervisor software to shift computing according to the availability of wind and solar power at various sites, and, says Martin Brooks, an independent research consultant working on the GreenStar Network, this works well enough to allow the network to handle even finicky applications like running a video server. The video, says Brooks, doesn’t skip even as the virtual machines hosting it are transferred, over an ultrafast fiber-optic network, between servers thousands of miles apart. “We have certainly had people consider [this project] outlandish, but we live it every day, so we don’t think that way,” he says.
Whether the Cambridge research will result in data centers in places as exotic as platforms in the middle of the Atlantic is anyone’s guess, says Hopper, who also admits that some of his visions for the project may be over the top. His colleague Sohan is less ambitious. “Sometimes when I talk to Hopper about this, I say that an easy way to bootstrap this project is to put a Sun modular data center in existing renewable energy sites.”
Sun already has a data center that fits in a single shipping container, notes Hopper. Getting one to a renewable energy plant is as simple as taking it there on a truck. Connecting it to the Internet, however, is another matter: the team’s models are based on the kind of high-speed fiber-optic networks that are available to academics but have yet to become economical for most commercial applications. Once they are, says Hopper, “we imagine putting photons into places that are godforsaken for every other reason except for generating energy.”
AI is here. Will you lead or follow?
Join us at EmTech Digital 2019.