Astrid Wynne | April 24, 2023
Image credit : Vector Stall/Shutterstock
Energy efficiency is key to cost savings, securing long-term energy supply and carbon reduction plans. All of these are at the forefront of business’ minds at the moment as well as being discussion points for governments worldwide.
There is growing awareness that reducing the amount of electricity used for each compute operation is vital if we are going to achieve energy efficient data centres. The Singapore Government are looking to put together standards on this, as is the European Commission through initiatives such as the Ecodesign Directive. They are looking for a system that will enable them to do this accurately for the wide variance of makes, models and generations of server on the market as well as the configurations of those machines. Likewise, enterprise customers are also asking the same questions. Because of the cost risks associated with electricity usage and potentially the security of supply, it has never been more important to make the best use of energy resources. When buying new equipment, procurement departments want to know what their best choices are for the precise systems they want to build. They need to understand which machines to buy, how much memory to build in, how many CPUs and of which type because these things make a huge difference to the amount of energy draw per operation.
Historically, people have looked to ENERGY STAR to provide a benchmark for sustainable servers. ENERGY STAR is the US government symbol for energy efficiency. It is well established, well-respected and well supported by the market. For these reasons, it has been built into standards and codes of practice worldwide. An ENERGY STAR rating is granted to all servers that use 30% less energy than conventional models. It means the machines meet minimum power supply efficiency levels, offer enabled power management features, and have high active energy state scores from SERT.
SERT (which stands for Server Efficiency Rating Tool) was developed by the SPECpower Committee, which includes representatives from AMD, Dell, HPE, IBM, Intel, Microsoft and the University of Würzburg. These organisations understand the complexity of IT hardware and which pertinent points to pull out for an overall score. SERT evaluates a server’s work capacity and power demand and gives this a score. The higher the score, the more efficient or productive the server is. The tool measures multiple utilisation levels across multiple “worklets” (or operation profiles). It aggregates these 4 utilization levels for 6 CPU worklets, 8 utilization intervals for one CPU worklet and at 2 intervals for 2 memory and 2 storage worklets into one overall score through a weighted geometric mean that allocates 65% to the CPU, 25% to memory and 10% to storage. Manufacturers submit their normalised SERT score to SPECpower, which lists these publicly.
ENERGY STAR’s tagline is “The Simple Choice for Saving Energy” and this is absolutely true. With the aid of SERT, it simplifies the bewildering array of variables associated with server energy efficiency into an easy-to-understand metric: pass or fail. (The SERT scores are listed in the ENERGY STAR product sheets, but it is the sticker that is most commonly used.) A grading system that was configuration specific and allowed for incremental improvements as a result of upgrades or reconfiguration would be much more useful.
Other products in ENERGY STAR include televisions, monitors, refrigerators, dryers and dishwashers. Whilst there will be wide variation within these product groups, the machines themselves are not subject to huge configuration changes. If you buy a washing machine, you will not get a choice of how many drums, pumps or heating elements it contains. The same is not true for servers, which are ordered more as bills of materials than off-the-shelf items. The sales process involves configuration of the machine, where the number of CPUs, amount and rank of memory are included depending on what the customer wants to do with it.
Varying configurations produce massive differences in performance and energy efficiency for servers. For example, the base configuration of a Gen9 HPE server is 13% less performant than the base configuration of a Gen10. Upgrade the CPU to a later iteration, and double the amount of Random Access Memory (RAM), and you make the Gen9 36% more performant than the Gen10 although it is nominally still a “Gen9”. There are in fact hundreds of variations on individual servers based on the configuration of key components and the variety of choice within these component groups. Despite this – or more accurately because of it – only one configuration will be submitted to SPECpower using SERT.
This in turn leads to a huge range of performance and energy efficiency scores between products that achieve an ENERGY STAR sticker. An example of this is the Dell R650 server, which has a SERT score of 80 based on a 48 core, 96 thread AMD CPU and 768GB of RAM and an ENERGY STAR listing. However, analyse the same server with different CPUs, total RAM and RAM ranking and you see a 40% difference in performance and energy efficiency between the best and the worst configurations.
Certified servers can have a SERT active state score of just over 9 or just shy of 60 and still be certified in the same way. Whilst these scores are listed on the ENERGY STAR site for comparison, they are not reflected in the certification itself. The low power scores are not listed on ENERGY STAR at all, which has implications for assessing the energy efficiency of servers installed for back up or resilince. All of this has implications for how effective ENERGY STAR certification is when assessing server estate efficiency for legislative purposes.
At Interact, we have been having a growing number of conversations with legislators, customers and manufacturers on building a grading system that can go beyond the work of ENERGY STAR and SERT scores. Customers provide five pieces of information (Make, Server Model, number of CPUs, CPU Model, and total RAM) for each server. We return a grading report (A+ to F) classifying your server estate from the most energy efficient to the least for each server.
It is a more comprehensive system than an efficiency certification and also more dynamic because it enables organisations to make the best buying choices at procurement stage as well as to improve the performance of existing estates. It makes it straight forward to identify low performers, drive beneficial decommissioning and identify servers for product life extension. We work from SPEC methodology because it is the best available (read our analysis of other benchmarks here). However, we build it out to encompass more iterations and more practical solutions for energy efficiency in the server estate. This provides data centres with the ability to accurately measure energy efficiency for the first time, a rigorous methodology the legislators are already calling for.