A thought experiment for a data centre containing 2,500 servers of the same make and model, each with a power draw between 47W (idle) and 272W (fully utilised), will ultimately reveal a wide span of power consumption needed (between 117kW and 680kW). Most data centre planners would lean towards the worst-case scenario of the 680kW plus additional power for the supporting infrastructure, which easily doubles the number. This approach realistically leads to a potentially significant oversizing of the supporting physical infrastructure.
Over the last decade manufacturers of data centre IT hardware have made tremendous progress in sub-optimising the efficiency of each of their subsystem devices, leaving us with a much better starting point for efficiency initiatives than ever before. However even the best-optimised device might be a negative contributor to the overall energy efficiency equation, if used improperly.
In the graphs below, it clearly shows that where performance and power usage was very linear with the CPU loads for the 2009-model, then the server manufacture have managed to de-couple the two for the 2015-model, achieving an optimal balance of CPU performance and power draw around 70-80% CPU load for the 2015-model versus 100% CPU load for the 2009 model.
Figure: Performance and corresponding power consumption comparison at various CPU loads between 2009 model server (right) and 2015 model server (left). Source: spec.org
Other servers will have different characteristics, i.e. peaking around 60-70% and seeing even worse performance around 90% to 100% CPU load. This means that a high level of awareness for each make, model and configuration of servers deployed in the data centre needs to be present, in order to strike the right balance between performance and energy efficiency.
With the functionality available today to initiate active power capping (the practice of limiting how much electrical power a server can consume) on servers based on Intel-based chipsets, data centre planners have a new option available when rightsizing the infrastructure without introducing unnecessary risk into the operation. For the example above with the 2015-model server, a power cap could be introduced at 70% CPU load, providing a design value of 191W load, resulting in a significant reduction to the overall requirements for the facility.
Another area where manufactures have made substantial progress over the years is in optimising the efficiency of data centre facility infrastructure equipment. A modern uninterrupted power supply (UPS), for instance, has a fairly linear efficiency (around 93-97% depending on make and model) when operating above 20-25% load. This means, that as long as the UPS is seeing a fair amount of load, there is not much to worry about in terms of running efficiently on the UPS component isolated.
Sign up for Computerworld eNewsletters.