* New measures of efficiency in data centers
What will be the most appropriate measurement of efficiency in the data center of the future?
In the past we explored new metrics for evaluating data centers, such as data center density (cycles per square foot), automation (admin staff per unit of data-center density), and storage density (storage utilization per square foot). These measures attempted to quantify both the rapid growth and the continuing concentration of compute power and storage in the limited spaces of the enterprise data center, managed by limited numbers of data center staff. They focus the attention on strictly internal parameters: space, staffing, compute cycles, storage utilization.
Since then, energy prices have continued to rise steadily, and awareness of both the financial and the environmental costs of enterprise computing has changed the way we view large data centers.
As we retool IT to be better stewards of both enterprise and natural resources, greening our data centers, other measures should now come to the fore as well, based on data centers’ interfaces with the environment. The basic means of interaction are energy consumption and heat production. So, other metrics merit consideration:
* Computational power efficiency, measured in cycles per kilowatt-hour, comparing compute activity to the energy required to enable it. For a true reflection of energy input, data center managers will measure actual inputs of electricity consumed rather than make calculations based on component ratings, and will measure all inputs, including power going to HVAC and lighting. (Feel free to substitute the SI units if you like metric units better; 1 kWh = 3.6 megajoules.)
* Computational heat efficiency, measured in cycles per BTU of heat produced. This compares computation to its main waste product, heat. This in some ways overlaps the measurement of energy efficiency, since the energy required to cool the data center must be included in the energy draw of the data center. However, it is a useful direct measurement of computational impact on the environment, especially in colder climates, where cooling can for several months in a year be a matter of circulating outside air rather than chilling hot inside air. The fact that cooling is achieved mainly by dumping heat directly out of the data center doesn’t obviate the fact that the heat is being produced. (Yes, you can also measure this in joules, 1 BTU = 1055 J, but remember that this is joules out, not joules in.)
To make these reasonable measures of actual work accomplished for energy in or heat out, IT must measure actual CPU utilization and factor out the idle clock cycles – they don’t contribute value, so they shouldn’t improve the efficiency rating.
Of course, the data center is about more than just computation, and the equivalent measures should be calculated for storage as well:
* Storage power efficiency, in terabytes per kilowatt-hour.
* Storage heat efficiency, in terabytes per BTU.
Again, IT must count only terabytes adding to the actual work accomplished: this won’t be all of them. An honest measure of energy consumption must count the empty disk space, but honest accounting of the utility provided by those spindles can’t. So these should be terabytes used per kilowatt-hour consumed or BTU thrown off.
Unless an enterprise has separated out storage from computation to the point where they are in isolated power and cooling domains, it will be measuring both storage and computational consumption or waste against the same energy inputs or heat outputs. This might make it more convenient to create a composite of some sort, simple or weighted, to reflect the shared inputs and outputs.
We should not forget the network, either. Data center networks can also draw a lot of power and create a lot of heat. A third pair of numbers, based on cumulative throughput on the data center network, could be added to get a more complete profile.
And of course, data center staffers generate heat too – over 350 BTU per hour on average! – so the more people have to be in the data center, the more heat they add to it and the less heat-efficient they make it. Efficient human operations, then, especially as measured by the automation metric discussed in the earlier column, will help reduce the waste-heat basis for heat efficiency calculations.
However we measure it, rising energy prices and climate change require that we begin to think about, quantify, and work to optimize the data center’s impact on the world – and on the corporate power bill.
