Two stories by James Glanz on power requirements of North American data centres were recently published in The New York Times. They drew my attention to the supposed environmental cost of unlimited anywhere, anytime computing and storage. It seemed that the Times wanted to cause a bit of a stir (it did) and challenge readers to think about a social good (anywhere, anytime computing and storage) additionally having a real environmental cost that needs to be lowered (I offer some thoughts on this below). The articles, which first appeared on 22 and 23 September 2012, conclude that data centres on average are not efficient consumers of power. Data centres reportedly applied on average only from six to 12 percent of the power utilized by the servers (including CPU, memory and storage) to perform computation. The rest of the power was used for computation readiness. Significant amounts of power were also used for cooling, and to keep the infrastructure available in the event of power failure on the public utility side (readiness of back-up power devices). The stories reported the presence worldwide of over three million data centres of various sizes, based on IDC data. The issues identified were:
- servers that have been left to run after the applications and users they supported have ceased to require them, so they have no computing role any longer, but simply wait, idle, for someone to request a service.
- servers that are occasionally required to serve peak demand, but which run idle most of the time because average demand does not require them
- servers that are dedicated to one application when multiple applications could be run on them
Power consumption is an environmental issue since power generation always has environmental cost. Waste of electricity and other environmental concerns were blamed on structural issues with electricity pricing, tax regimes, risk management (peak capacity planning, power-supply redundancy and fear of failure on these fronts).
- From a pricing point of view, utilities are motivated to discount pricing for heavy users with continuous draw on the grid. Bulk buyers needing similar wattage 7 x 24 get large discounts for their power and even sometimes face penalties for under-consuming their predicted demand. Therefore, under existing contracts, a data centre might see no cost benefit from restructuring its operations to scale power consumption up and down more directly with demand, even if it were technically possible to do so.
- In the same vein, data centres pay relatively little for power compared to the perceived cost of failing to satisfy their customers. End-users are impatient and expect immediate response whenever they make a request. In a competitive marketplace for responsive cloud services, delays can cost good will and business.
- In addition to wasting grid power, a second environmental issue relates to backup systems for grid power. Large numbers of diesel generators stand beside data centres, ready to replace grid power should it fail. They must be routinely started and stopped, creating significant amounts of air pollution during their maintenance cycles.
Criticism of the articles has included claims that the stories 1) gave too little credit for recent improvements in data centre architecture and related reductions in power consumption, 2) oversimplified the issues that drive power consumption, 3) mixed different issues (grid power consumption versus backup generator-related air pollution). Some writers also 4) questioned the statistical reliability of the survey data cited in support of the power-consumption data, 5) suggested that there were some important questions left unasked. One of the criticisms that seemed to me fair was the lack of space allocated to the question of what is a reasonable power-use target. In the first article, Glanz described how the National Energy Research Scientific Computing Center was able to drive utilization up over 90 percent by queuing and scheduling large jobs at the Lawrence Berkeley National Laboratory, while acknowledging that not all computing requirements can be queued and scheduled. Technology to power down servers when not required was cited, but not details that would allow us to know whether it was suitable for all commercial data centres or its limitations. The author combined information about presumed efficient and inefficient data centres to create a sad picture of what might be “waste”, but it is hard to say from the evidence presented how much waste and what the best remedies would be. Is an independent analysis even possible? The general level of secrecy required by large and highly competitive players in the data centre business can make analysis difficult, given all the possible variables and lack of disclosure. Are data centres efficient users of electricity? Perhaps many are, but it stretches credulity to suggest that assigning from 88-94 percent of electrical power on average to non-data-processing related functions is a best-case scenario for a data-processing enterprise. We know that batch processing can dramatically improve things, and that some data centres fail to decommission or repurpose servers when they ought to. One would think that we could do better than this, on average. A data centre may be like a highway that, once built, cannot be rolled up when no-one is using it, but that analogy applies to the physical infrastructure, not directly to the power consumption. One would hope that the processing components, such as servers and storage, could be readiness-scaled in response to predictable demand, in the way that electricity producers take capacity offline and bring it back based on weather forecasts, cyclical use agreements with customers, and so on. And one hopes that as data centre power consumption best practices become more pervasive, the end result will be positive change in those striking metrics.