Methodology: 4 Steps
In a recent blog post, I presented a 4-step approach to compare data center TCO to costs involved in consuming resources from an external cloud provider. Here it is again:
- You need a model to calculate the TCO of your infrastructure (or the part of your infrastructure that you are planning to “cloudsource”)
- How do you deal with under- and over-utilization? What are the risks and costs involved if something crashes (over-utilization)? What are the costs of running idle servers (under-utilization)?
- What would be a comparable amount of cloud resources (provided there is a utility computing model involved)? How do you compare resources?
- Calculate!
Step 1: Data Center TCO
First I must determine the TCO of my IT infrastructure. Since I do not have insight into real enterprise data to do this, I use a great model provided by Jonathan Koomey from Stanford: “A Simple Model for Determining the True Total Cost of Ownership for Data Centers” (Source).
My assumptions: I have a High Performance Computing (HPC) data center with around 6000 IBM X-3550 1U servers with 8 GB RAM and two dual core 2.66 GHz Intel processors on an electrically active area of 20,000 square feet (with another 20,000 square feet for cooling systems and energy infrastructure). The facility goes on-line when it is completely built and there are no costs for decomissioning, equipment disposal, etc.
CapEx sums up to around $100 M, with $29 M for IT and $72 M for other expenses besides IT, i.e. racks, cables, external hardwire connections, etc. Assuming a lifetime of 3 years for IT equipment and 15 years for the facility, results in average annualized capital costs of ca. $25 M per year.
Step 2: Utilization
The most important aspect when comparing costs of a data center with a flexible cloud computing pricing model certainly is the utilization ratio of the data center. Owners of data centers usually struggle with two contradictory objectives. On the one side, the data center CapEx should “pay off”, which means: increase server utilization as much as possible. On the other side, it might be a good idea to save some server capacity for future projects and emergency cases.
This is something that is hard to measure in $ because it depends on the risks and costs involved in case that your service goes down.
Step 3: Compare resource usage
Now I must transform the data center resources into a utility computing model that makes them comparable to a cloud offering (in this case Amazon EC2 and S3).
This is perhaps the weakest part of my calculation since I do not know enough about bare metal server technology to compare physical resources from the data center TCO model in step 1 with servers in the Amazon cloud.
Step 4: Results
For my HPC data center the results show that up to a utilization ratio of 70% it would be cheaper to do things in the cloud. A highly efficient data center with an average utilization beyond 70%, on the other hand, would beat the cloud. The open question is whether or not (or to which degree) you can scale your own infrastructure.
