Data Centers

A holistic metric pits energy demand vs. supply to improve data center efficiency

Resource Allocation Index (RAI) is one researcher's answer to realistically measure and model data-center energy usage and efficiency.


Acronyms abound in IT, and data center discourse is certainly no exception. In this article, I'll discuss PUE (Power Usage Effectiveness), OPEX (operating expenses), and RAI (Resource Allocation Index).

In my recent article about PUE, I wrote that data-center designers are less than satisfied with PUE as an accurate measure of how efficiently a data center uses electricity. I also reported that several new data centers are measuring PUE in real time — Facebook's Prineville, Oregon data center is the example I used. Figure A advertises the Prineville data center as having a real-time PUE of 1.05. The efficiency experts are less resistant to real-time PUE, but there's a cadre who feel PUE is the wrong metric because of the variables being measured.

Figure A

 Image: Facebook

PUE is the wrong metric

Rajat Ghosh
One individual belonging to that group is Rajat Ghosh, a Postdoctoral Fellow at the CEETHERM (Consortium for Energy Efficient Thermal Management) Laboratory located on the Georgia Institute of Technology campus. The laboratory is an 1,100 square-foot simulated data center used by researchers to study thermal and energy management of electronics in data centers.

In a Data Center Knowledge article, Ghosh asserts that in order to optimize OPEX, data center operators should manage resource consumption by adopting the "usage-based" pricing model — i.e., pricing a service or item based on its consumption or usage, rather than a flat rate for a given service or period of time. This approach suggests matching the data center's resource supply side and demand side as closely as possible. Figure B depicts the resource utilization value chain.

Figure B


The demand side could be anything from hosting a banking website to an online merchandising website such as Amazon. It's important to understand that incoming traffic places demands on the data center, and that the work required by the data center to comply with the incoming demands requires a certain amount of electricity that can be calculated.

Next, Ghosh explained the resource supply side:

"Although the data center's electricity should be primarily consumed by its IT equipment, a few exhaustive surveys indicate that on average 35-45 percent of a data center's electricity is consumed by its cooling hardware."

That is the relationship measured by PUE. It's useful, but Ghosh believes data center managers need "A holistic metric that can encompass the entire resource utilization value chain (both supply and demand) by a singular metric."

RAI is the right metric

Ghosh calls this metric RAI, where RAI equals Normalized Resource Supply divided by Normalized Resource Demand. Or put in simpler terms, RAI measures how much electricity is required by a data center in order to serve one request.

He then tackles how the supply and demand measurements are obtained. The demand side is measured using socket-level logs or the data center's preferred method of tracking inbound traffic. Supply side measurements are obtained from UPS/PDU logs. Because demand and power requirements change constantly, RAI uses three ranges:

  • Under-provisioning: This may seem the most economical, as the data center spends less money in order to meet demand, but under-provisioning power could also lead to degraded service.
  • Optimal-provisioning: This window is the sweet spot; the amount of electricity used is close to the amount of electricity required.
  • Over-provisioning: Too much electricity is being used for the current demand.

RAI allows data center managers to be proactive

Since PUE does not consider demand, it's possible for the PUE to look good, even though no billable work is passing through the data center. RAI considers the workload (demand), which means the amount of electricity can be customized to meet the current demand. Figure C is a representation of a data center's resource allocation. The graph depicts the workload one would expect, except for the unusual activity at 2 AM.

Figure C


Figure D displays the supply, demand, and RAI values based on the above graph.The supply of electricity was cut in half for non-business hours, and the data center was optimally provisioned except for the 2 AM activity.

Figure D


Any data center can reduce power during non-business hours, but without RAI there is no way of determining whether the power down is affecting the demand side.

Ghosh ended his article by noting that RAI can be used:

  • To compare resource allocation performances of different data centers, and
  • To assess how a data center's resource-provisioning status changes with time.


Information is my field...Writing is my passion...Coupling the two is my mission.

Editor's Picks