Developing a Cost Model for Cloud Applications

Note - please pay attention to the date of this post. As much as I attempt to make the information below accurate, the nature of distributed computing means that components, units and pricing will change over time. The definitive costs for Microsoft Windows Azure and SQL Azure are located here, and are more accurate than anything you will see in this post: 

When writing software that is run on a Platform-as-a-Service (PaaS) offering like Windows Azure / SQL Azure, one of the questions you must answer is how much the system will cost. I will not discuss the comparisons between on-premise costs (which are nigh impossible to calculate accurately) versus cloud costs, but instead focus on creating a general model for estimating costs for a given application.

You should be aware that there are (at this writing) two billing mechanisms for Windows and SQL Azure: “Pay-as-you-go” or consumption, and “Subscription” or commitment. Conceptually, you can consider the former a pay-as-you-go cell phone plan, where you pay by the unit used (at a slightly higher rate) and the latter as a standard cell phone plan where you commit to a contract and thus pay lower rates. In this post I’ll stick with the pay-as-you-go mechanism for simplicity, which should be the maximum cost you would pay. From there you may be able to get a lower cost if you use the other mechanism. In any case, the model you create should hold.

Developing a good cost model is essential. As a developer or architect, you’ll most certainly be asked how much something will cost, and you need to have a reliable way to estimate that. Businesses and Organizations have been used to paying for servers, software licenses, and other infrastructure as an up-front cost, and power, people to the systems and so on as an ongoing (and sometimes not factored) cost. When presented with a new paradigm like distributed computing, they may not understand the true cost/value proposition, and that’s where the architect and developer can guide the conversation to make a choice based on features of the application versus the true costs.

The two big buckets of use-types for these applications are customer-based and steady-state. In the customer-based use type, each successful use of the program results in a sale or income for your organization. Perhaps you’ve written an application that provides the spot-price of foo, and your customer pays for the use of that application. In that case, once you’ve estimated your cost for a successful traversal of the application, you can build that into the price you charge the user. It’s a standard restaurant model, where the price of the meal is determined by the cost of making it, plus any profit you can make.

In the second use-type, the application will be used by a more-or-less constant number of processes or users and no direct revenue is attached to the system. A typical example is a customer-tracking system used by the employees within your company. In this case, the cost model is often created “in reverse” - meaning that you pilot the application, monitor the use (and costs) and that cost is held steady. This is where the comparison with an on-premise system becomes necessary, even though it is more difficult to estimate those on-premise true costs. For instance, do you know exactly how much cost the air conditioning is because you have a team of system administrators? This may sound trivial, but that, along with the insurance for the building, the wiring, and every other part of the system is in fact a cost to the business.

There are three primary methods that I’ve been successful with in estimating the cost. None are perfect, all are demand-driven. The general process is to lay out a matrix of:

  • components
  • units
  • cost per unit

and then multiply that times the usage of the system, based on which components you use in the program. That sounds a bit simplistic, but using those metrics in a calculation becomes more detailed. In all of the methods that follow, you need to know your application. The components for a PaaS include computing instances, storage, transactions, bandwidth and in the case of SQL Azure, database size. In most cases, architects start with the first model and progress through the other methods to gain accuracy.

Simple Estimation

The simplest way to calculate costs is to architect the application (even UML or on-paper, no coding involved) and then estimate which of the components you’ll use, and how much of each will be used. Microsoft provides two tools to do this - one is a simple slider-application located here: 


The other is a tool you download to create an “Return on Investment” (ROI) spreadsheet, which has the advantage of leading you through various questions to estimate what you plan to use, located here: 


You can also just create a spreadsheet yourself with a structure like this:

Program Element Azure Component Unit of Measure Cost Per Unit Estimated Use of Component Total Cost Per Component Cumulative Cost

Of course, the consideration with this model is that it is difficult to predict a system that is not running or hasn’t even been developed. Which brings us to the next model type.

Measure and Project

A more accurate model is to actually write the code for the application, using the Software Development Kit (SDK) which can run entirely disconnected from Azure. The code should be instrumented to estimate the use of the application components, logging to a local file on the development system. A series of unit and integration tests should be run, which will create load on the test system.

You can use standard development concepts to track this usage, and even use Windows Performance Monitor counters. The best place to start with this method is to use the Windows Azure Diagnostics subsystem in your code, which you can read more about here: This set of API’s greatly simplifies tracking the application, and in fact you can use this information for more than just a cost model.

After you have the tracking logs, you can plug the numbers into ay of the tools above, which should give a representative cost or in some cases a unit cost.

The consideration with this model is that the SDK fabric is not a one-to-one comparison with performance on the actual Windows Azure fabric. Those differences are usually smaller, but they do need to be considered. Also, you may not be able to accurately predict the load on the system, which might lead to an architectural change, which changes the model. This leads us to the next, most accurate method for a cost model.

Sample and Estimate

Using standard statistical and other predictive math, once the application is deployed you will get a bill each month from Microsoft for your Azure usage. The bill is quite detailed, and you can export the data from it to do analysis, and using methods like regression and so on project out into the future what the costs will be. I normally advise that the architect also extrapolate a unit cost from those metrics as well. This is the information that should be reported back to the executives that pay the bills: the past cost, future projected costs, and unit cost “per click” or “per transaction”, as your case warrants.

The challenge here is in the model itself - statistical methods are not foolproof, and the larger the sample (in this case I recommend the entire population, not a smaller sample) is key.

References and Tools



Other Tools:

Comments (0)

Skip to main content