One paper comes from Microsoft, called The Cost of a Cloud: Research Problems in Data Center Networks. It focuses on the general cost structures of data centers and discusses how they can be reduced. The second is the work of a single author named Steve Denegri and is called Cloud Computing Dominance Through Renewable Energy; it was originally published by Virtual Strategy Magazine. Virtual Strategy has very kindly arranged to make the paper available at its website.
The Microsoft paper describes the challenges of reducing overall costs structures of data centers and discusses various strategies, including migrating redundancy and robustness measures (e.g., UPS) from individual data centers to the agglomeration of data centers. With this approach, less investment is required, since costs associated with hardening individual data centers can be foregone. Another area the paper addresses is the overall power draw of data centers; while not necessarily proposing specific technologies to reduce power use (save running the center at a higher overall temperature to reduce AC costs), it does recommend finding cost incentives to motivate system users to spread demand to a more even use pattern. This would have the effect of reducing peak demand and therefore reduce overall power consumption. The paper contains a table on Page 1 identifying the various components of data center costs; servers are the number one cost at 45%, but "power distribution and cooling and electrical utility costs" are numbers two and three, summing to 40%.
The second paper, as indicated by its title, focuses on the energy component of data centers. The author, who I had the chance to speak with after reading his paper, is a former investment banking financial analyst; his facility for research and digging into numbers certainly shows in the paper. He has done a lot of digging into the cost structures of data centers; specifically, he decomposed the cost structure of a huge Microsoft data center (on the order of 100K servers in a single facility) in San Antonio, Texas. His paper quotes a Microsoft source (uncited, unfortunately) that energy comprises 82% of the total cost of running a data center.
Denegri digs into the deal that Microsoft struck with the City of San Antonio for property tax relief as well as with the local power company, CPS Energy. From his research, he concludes that Microsoft obtains a 22% discount on its property taxes (this is based on a discount extended to companies that make very large real property investments). With regard to power, Microsoft obtains a 45% discount, based on using enough electricity to fall into a "Super Large Power" user category. This advantage may be enhanced in the future, Denegri states, because of potential taxes to be applied to carbon-based power. Microsoft may be able to strike an arrangement to have its data center powered by clean energy sources (e.g., wind or solar), thereby avoiding these additional taxes.
One cost factor touched only briefly addressed in both papers is labor costs. Both note that these large data centers tend to run leanly staffed; indeed, the Microsoft paper specifically notes that one factor in reducing costs is the high level of operations automation in these cloud centers. In fact, it states that one of the largest source of problems in data centers is manual management, which introduces breakages. It goes on to say that another factor in lean operations costs is hardware standardization. In contrast to corporate data centers, in which the servers tend to be procured on a project-by-project basis, which introduces heterogeneity, cloud centers focus on homogeneity in hardware.
Denegri notes that in the Property Tax agreement Microsoft struck with San Antonio, it received the most favorable tax break despite not qualifying on the minimum number of jobs required: instead of at least 500, Microsoft's facility will employ an estimated 75 people. That works out to somewhere in the neighborhood of each employee responsible for around about 1000 servers. To assess what advantage a cloud provider might have versus an internal data center, two things must be recognized:
1. There are no hard numbers on how many servers an operations person can manage in a traditional corporate data center. In digging around, I found estimates ranging from 75 to 200.
2. It must be acknowledged that the a cloud operations person and a data center operations person is not a directly comparable position. This is because operations in a traditional large data center encompasses the entire hardware and software stack, whereas in a cloud environment, operations is responsible for the hardware and some amount of the software stack, but definitely not the application itself.
Nevertheless, it seems clear that cloud centers must, through their homogeneity and standardization, achieve a significant efficiency gain for operations personnel. Perhaps it could be placed at 50%, though I think that understates it.
With these two papers in hand, let's perform an interesting experiment. Let's attempt to estimate the total cost advantage one of these cloud data centers has versus a traditional corporate data center. The individual factor cost advantages estimated are:
1. Power: 45%; this comes from the Denegri paper. 2. Property Taxes: 22%; this also comes from the Denegri paper. 3. Operations Staff Costs: 50%; I've just worked out this estimate by comparing the headcount numbers from the Denegri paper to research I performed on the Internet.
The next question is, what proportion of the total cost do each of these factors represent? Here is a swag:
Power: 40%; this comes from the Microsoft paper, please note that the Denegri paper quotes Microsoft (a different Microsoft source from the authors of the paper, presumably) as estimating power at 82% of total data center cost. By using the lower Microsoft paper estimate, we are reducing the total cloud center cost advantage vis a vis traditional data center.
Property Taxes: 20%. In the Denegri estimates, he calculates total costs, and property tax comes to around 20% of the total.
Operations Staff Costs: 40%. This is the remainder of total cost left to be assigned after removing Power and Property Taxes; however, it accords pretty well with the observation in the Microsoft paper that operations is a very large percentage of total data center cost.
To calculate the overall cost advantage of a cloud computing data center (a Microsoft one, in this example) versus the traditional data center, we multiply individual cost factor cost advantage by the proportion of total cost the factor represents. The Weighted Factor Advantages are then summed to determine the total cost advantage of a cloud computing data center.
We can see that, based on the estimates from the papers and a calculation of operations staff costs, Microsoft can achieve an approximately 42% lower cost than a traditional data center. I believe this actually understates the total cost advantage, but this is a fairly conservative assessment. Also, the advantage will vary from cloud data center to cloud data center, depending upon property tax concessions and the type of power available at a given location.
In examining these figures, I'm reminded of the move to mass production started by Henry Ford. He assiduously applied automation techniques to auto manufacture, with the result that the cost of producing a car dropped by 80 to 90 percent. Rooted in historical antecedent, corporate data centers have remained rooted in expensive, ponderous, and fragile manual techniques. Cloud computing has applied the industrial automation paradigm to data centers and is evincing the same kind of cost advantages Ford's manufacturing techniques brought to the auto industry. When Ford started his automation campaign, there were several hundred auto manufacturers; 25 years later there were probably only a dozen or so. More than 90% of them had vanished, unable to afford the capital to invest in mass production and therefore uncompetitive on pricing. Perhaps there's a lesson there for all of us.
Bernard Golden is CEO of consulting firm HyperStratus, which specializes in virtualization, cloud computing and related issues. He is also the author of "Virtualization for Dummies," the best-selling book on virtualization to date.