Cloud, Colocation, and Data Center Real Estate Analysis
With the maturation of the IT infrastructure business, solution sets for any given need have multiplied and produced an alphabet soup of viable alternatives. IaaS, Cloud, Colo, PaaS, Container, etc., all serve essentially the same need for infrastructure services. Processor level advances have solved many problems, and now bottlenecks have emerged elsewhere in the solution stack, namely the supply of ready-to-occupy (data center) infrastructure capacity. This multiplicity of solutions is creating scope creep in the infrastructure decision making process, resulting in analysis paralysis in the best case and millions of dollars spent egregiously in the worst case. This article will share insight into how to create transparent and fact-based comparisons that will i) simplify the decision making process, ii) allow one to avoid common mistakes, and iii) achieve a better fit between infrastructure requirements and an associated solution.
Whether itÃ¢â‚¬â„¢s contracted or developed internally, a pre-requisite to a large scale deployment is good engineering and project management talent. And what is good engineering? If given enough money, anyone could build a highly efficient cloud, colo or enterprise scale data center. It may take more than one try, and it may take several years, but it can be done. Today, many IT staffers with little experience in power, telecom, finance, law, or facilities are doing exactly that, directing investments worth millions of dollars to little effect. Good engineering and project management will achieve the desired result on the first attempt, on schedule, and to an attractive budget.
Secondly, whether simple or complex, it’s important to create financial models, know the determinacy of inputs (risk assessments), and use numerical benchmarks, not subjective assessments, to drive the project development and management activities of your team. Today, the dollar per kilowatt metric ($ / kW-month), whether on a cash or NPV basis, is the gold standard for comparing large-scale infrastructure alternatives.
In addition, these comparisons are rooted in investment analysis and asset management, which are mature disciplines. Cloud may be a new (and now over-hyped) solution, but the need for and application of flexÂ capacity in infrastructure intensive industries has existed for many years and can be numerically assessed via cost and efficiency. The addition of a professional with asset management experience from a mature enterprise can add significant value to any firm that has scale investments in the data center space.
The emergence of cloud is evidence of a maturing industry moving toward greater efficiency and closely parallels a similarly asset intensive industry, namely utilities. This similarity can serve as a useful guide for first-pass due-diligence in determining which solution, whether cloud, colo, or real estate, will best suit your needs. Below is a graphical depiction of the application of both utility and IT solutions. On the left are listed three basic categories of power plants, ranked by their load following capabilities, and on the right are the corresponding IT solutions. Technically, each time a light is turned off or on, a small unit of power must be added or subtracted from the power grid. Carbon aside, coal fired power plants are highly efficient and produce cheap electrical power, but they take hours to start and large dollars to build (e.g. a large, fixed-cost asset), making it both uneconomic and infeasible to follow a load. In order to maximize the benefits of their low-cost output, they have to be operated at 100% capacity (or more) at all times. The multi-megawatt data center is the closest cousin of the base load power plant, as it is a large, fixed-cost asset with limited load following capability but attractive, low unit cost. Based on transactions and market searches conducted by WiredRE nationally, the TCO of a 1 to 3 MW wholesale, fully operated data center is approximately $200 / kW, including power and cooling, assuming a 75% load factor, 1.3 PUE, and $0.06 cents / kW (excluding fit-out).
This cost is far more economical than that of the near just-in-time, flex solution of colocation. Metrics from publicly traded, carrier-neutral providers such as Equinix are useful benchmarks, as these companies are nearly a pure-play real estate firms. On average, their customers pay over $600 per kW or three times the wholesale rate.
Peak or just-in-time (“JIT”Â) capacity may be third in this list, but that’s not because it’s unimportant. The importance of load following in IT infrastructure grows with complexity, scale and money. Consider that according to researchers at Google, the head room for their large-scale data centers is 40%, owing to the poor load following capabilities of fixed data centers and over-provisioning [Fan 2007]. As shown above, wholesale data centers are much more economical than colocation, so managing to 100% utilization via flex or JIT solutions can translate into real money. Utilities invest in peaking plants that produce electricity just-in-time, for consumption, or ramp down in the event of loss of demand, minimizing overhead and targeting this 100% utilization in their most capital intensive assets. At the end of the day, and after all the hype is gone, the reasons for selecting cloud computing are nearly identical to a utility’s goals for peaking plants, namely i) improved manageability, ii) JIT provisioning, and iii) lower TCO [Forrester 2009].
If utilities aren’t intuitive, then the now numerous examples from both forward thinking and traditionally conservative organizations may be used as guides for cloud implementations with the goal of greater efficiency (e.g. load following). It may also help to think of cloud in terms of managing the indeterminate (e.g. risk) around time and capacity. Moreover, it’s critical to assess the value and new capabilities that access to very short term but very high peak capacity can create. The Indy500.com example listed below, where cyclical and short-term capacity needs can be met and then turned off, is efficiency defined.
With a few additional metrics and tools in hand, financial and sensitivity (risk) analysis can easily quantify efficiency and make deciding between cloud, dedicated hosting, colocation, or data center real estate simple. For the purposes of this article, we’ve used the simple financial model available from Amazon (see http://aws.amazon.com/economics/). The basic scenario that comes with the model assumes 300 low-end servers, at 75% annual utilization (percent of hours not load). For our purposes, we are going to assume 24×7 operation (100% utilization), and we are going to assume 300 low-end servers and compare them with 300 high-end (High-Memory Quadruple Extra Large) servers. Removing assumptions for network usage, the total cost is $47 / month for on-demandÂ or $27 / month for three-year reservedÂ individual server instances.
These are clearly attractive figures that demonstrate the operational leverage that Amazon has achieved in its infrastructure. However, changing the assumptions to 300 high-end instances shows a total cost of $1,748 / month for on-demandÂ or $884 / month for three-year reserved individual server instances. We benchmarked these figures against the dedicated server costs of Peak Web Hosting. Peak Web charge $500 / month for a dedicated server of comparable compute capacity, including all necessary network infrastructure and other peripheral infrastructure, which means at a monthly total hourly utilization (not load) of just under 30% (209 hours or less than 9 days in a month), it’s only break even with dedicated hosting, which means, 300 servers or not, it’s probably preferable to have a dedicated resource that can serve many uses, unless you can really take advantage of the ability to scale quickly and then turn off instances, or when the estimates of those boundaries are highly indeterminate.
Comparing cloud to colocation costs with the same model and above scenarios yields similar results, with the low end favoring cloud and the higher end favoring colocation. The model yields $98 / server for the lower end scenario and $901 / server for the high end.
Any of these comparisons are highly dependent upon staffing assumptions, which is one of the weaknesses of the Amazon model, as well as performance considerations, and of course feasibility issues related to regulatory compliance or data base dependencies and VM support. How many more dedicated servers can an admin manage in a colo, or dedicated servers with Peak Web or Rackspace, versus server instances with cloud? According to CIO Magazine, enterprises like Bechtel and others only manage 100 servers, while firms like Microsoft and Google manage several thousand per admin. If tapping the cloud allows you to manage an extra 200 servers, then the answer changes rapidly at the low end. But then again, it’s possible to run your own VMs as well.
One question leads to another, but the approach is the same.
[Fan 2007] X. Fan, W.-D. Weber, and L. A. Barroso. Power Provisioning for a Warehouse-Sized Computer. In Proceedings of the Thirty Fourth Annual International Symposium on Computer Architecture, 2007.
[Forrester 2009] James Staten, Senior Analyst. Cloud Realities Presentation. Wired Real Estate Group, Data Center Dilemma Event Series, 2009.