Figure 1. The IDC provides data on the energy used by servers in the U.S. and worldwide.

In economics, the scarcity of resources drives the cost of those resources, which results in more efficient use of those resources. When gasoline prices go up, consumers demand more efficient vehicles to offset those increasing costs. Conversely when fuel prices go down, the demand would go up substantially. This latter notion is called Jevons’ Paradox.

Similarly, the well-known phenomenon called Moore’s law has substantially reduced the cost of computation year over year, driving demand up. This cost reduction is completely driven by efficiency improvements in electronics over the past half-century, improvements unmatched by any other industry. This is the most important reason for the significant demand for increases in the number of servers in the data center and more importantly the increase in server power footprint worldwide. So the real question is how would the further efficiency improvements suggested by the Environmental Protection Agency (EPA) in its 2007 report to Congress affect these trends?



Figure 2. EPA projection of energy footprint in the U.S. with increasing efficiency.

The EPA says that driving efficiency into the data center culture can reduce the growth of power demand in data centers. The EPA suggests changes that it believes could even reduce the total load; However, Jevons’ Paradox suggests that increased energy efficiency may cause demand to increase even more rapidly in the future. While clearly the efficient use of resources should be the goal of all enterprises, economists and enterprises must hold a dialog around the possibility of unintended consequences of approaches designed to achieve greater efficiency.


Figure 3. Raw performance and performance/watt increase in a example server.

Background

Over the past couple of years, power consumed in the data center by IT equipment has become a center of attention for many businesses as well as government agencies all over the world. Jon Koomey’s report (Estimating Total Power Consumption by Servers in the U.S. and the World, http://enterprise.amd.com/us-en/AMD-Business/Technology-Home/Power-Management.aspx) shows why. In five years, energy consumption in data centers both in the US and globally has doubled (see figure 1). In 2005 data centers consumed 1.2 percent of electricity produced in the US and 0.8 percent of the world’s production, both doubling since 2000.

While this growth doesn’t seem like much, the EPA is concerned not about the amount of power but rather the rate of growth of the power footprint. In its 2007 report, the EPA projected that if the industry doesn’t improve efficiency, power consumed by data centers will double again in the next five years. Figure 2, which is taken directlyfrom the report, shows this quite clearly; by 2011, the annual electricity use will be over 120 billion kWh/year, more than double that used in 2006.

In addition, the figure shows the energy savings that would be achieved by four different energy-efficient scenarios. The most aggressive of these scenarios actually achieves reduced electricity demands (“State of the Art Scenario”).

Figure 4. Annual amortized costs in the data center for a 1U server.

The Economics of Efficiency

Figure 2 clearly shows the EPA’s position; aggressive efficiency strategies and best practices can mitigate the increasing footprint of servers. On the surface this makes sense, but one needs to look at what is driving the growth in server power in the first place. Figure 3, which appeared in the February 2007 issue of Electronics Cooling (In the Data Center, Power and Cooling Costs More than the IT EquipmentIt Supports) shows the performance and efficiency of a server that was tracked over an eight-year period.

During that period, raw performance of the server increased 75 times, and the performance per watt increased 16 times. No other technology in the history of the industrialized world has exhibited this kind of efficiency gain in such a short period. This would be the equivalent of a car that got 20 miles per gallon (mpg) eight years ago getting 320 mpg today.

An economist might use figure 3 to consider computation as a resource, especially given the assumption that the cost of the computer stayed flat during the eight-year period (if anything its cost actually has gone down). In this case, the actual cost of computation would have decreased by 16 times over the period. In other words, the cost of running the identical business application would be one half of what it was two years prior. These economics drive the insatiable demand for computation resources. As the cost of computation drops, demand for it increases because the reduced cost improves the ROI story of more and more business applications. By way of explanation, if the price of gasoline dropped from $3.20/gal to $0.20/gal in eight years, people would drive more and might even find ways to run refrigerators and other plug loads on gasoline-powered systems, driving demand up.

Figure 5. Jevons’ Paradox illustrated

It is fair to assume that efficiency and cost reduction in computation is one and the same. Figure 4 shows how the cost of supporting a server has gone up with server power, so much so that the cost of computation is heavily driven by the data center infrastructure. In fact the cost of the infrastructure is actually greater than the cost of the server since the cost of the infrastructure scales with power. Any improvement to server efficiency not only reduces the amount of energy required but also reduces the amount of infrastructure and thus the cost of compute, driving the demand for computation up.

Jevons’ Paradox informs a new thesis about the EPA’s energy-efficiency proposals: Increased efficiency in the IT environment will actually increase the demand for servers and could actually increase the power footprint. Unlike the EPA report, which assumes that demand stays the same with efficiency, Jevons’ Paradox suggests that demand will increase. The cost decrease brought on by efficiency improvements actually does the reverse of what is expected.



Figure 6. Jevons’ Paradox applied to EPA chart.

Jevons' Paradox

Wikipedia describes this Paradox as follows:
 "In economics, the Jevons’ Paradox is an observation made by William Stanley Jevons, that as technological improvements increase the efficiency with which a resource is used, total consumption of that resource may increase, rather than decrease. It is historically called the Jevons’ Paradox as it ran counter to Jevons’ intuition. However, the situation is well understood in modern economics. In addition to reducing the amount needed for a given output, improved efficiency lowers the cost of using a resource – which increases demand. Overall resource use increases or decreases depending on which effect predominates.”

Essentially, if efficiency does lower the cost of computing, it is actually a shift down in the supply curve of the compute resource.as shown in figure 5. Note that the actual quantity increases from Q1 to Q2 while the price drops from P1 to P2.

Bring Jevons’ Paradox together with Moore’s law and the result is what we see today in the IT industry, an insatiable demand for continually cheaper compute resources. Over the past couple of years, the argument has been that efficiency and total cost of ownership (TCO) are aligned, but the argument can also be made that improved efficiency will reduce cost even faster then what is resulting from Moore’s law. To illustrate this, figure 6 demonstrates what might be expected because of efficiency improvements when demand changes are considered. Note that this is exactly contrary to the EPA’s thesis.

Conclusion

While driving efficiency is clearly the right thing to do, industry and government agencies must consider a more holistic view on the cause and effect in the IT industry. It is important to understand what the drivers are to the perceived problem of IT power consumption. To date, this “problem” has been tackled as a technical problem and organizations such as the EPA, Green Grid, Climate Savers are doing the right thing to solve the technical problem.

There is no doubt that there are inefficiencies in the data center and in IT equipment, but perhaps it is exactly these inefficiencies that have curbed what the power growth could have been. In addition, it may be that IT is displacing other high-carbon emission industries. Drawing economists into the discussion alongside technologists could lead to a broad discussion of the larger issues and holistic solutions. The hope is that economists will be just as passionately involved in solving this problem as the technologist over the past couple of years.

Finally, the question is always asked, what factors will dampen the effect of increasing IT power footprint?

Moore’s law runs into physical limitations due to manufacturing processes with Silicon such that performance increases are no longer possible. This is not likely in the short term.

Infrastructure costs rise faster than the improvements gained by Moore’s law. Also unlikely in the short term.

Energy resources become and scarce and prices rapidly increase. This is really the most unpredictable globally. Energy costs will go up but as long as they go up less than the increase in performance the power footprint is likely to grow. This is the problem for economists to solve.