DIGITALEUROPE Analysis on ErP Lot 9 Idle Power Framework and Allowances for Enterprise Servers
DIGITALEUROPE Analysis on ErP Lot 9 Idle Power Framework and Allowances for Enterprise Servers
DIGITALEUROPE has serious concerns with the proposed base idle state power allowances and the additional idle power allowances for extra components. These allowances come directly from an initial draft proposal made by EPA for ENERGY STAR server specification which have no strong foundation prior to wider stakeholder detailed comments and analysis including that of Industry. As proposed in the Draft Lot 9 working document, the idle power threshold proposal will drive an estimated 10-30% net increase in data center energy use and consumption (Section 4).
DIGITALEUROPE, in conjunction with the Green Grid, has completed an analysis of the relationships between idle power, server active efficiency, deployed power and system launch year. The Green Grid SERT Analysis Working Group (TGG) has a data set of 71 one and two processor socket rack server products, with each server product having reported SERT data for three to five configurations. 262 separate configurations have measured idle power values and SERT metric scores. The data set also has 26 blade and 6 resilient two socket product families. The data analysis shows that:
1. Average server idle power has stayed largely flat by configuration over the past 6 years (since 2011) while average performance has steadily increased with each new generation of products. Server manufacturers have been increasing the performance capabilities of server products while maintaining a relatively constant power envelope (Section 1).
2. Higher performance servers, using more capable processors as measured by core count and frequency, will result in lower power consumption in the data center under both active and idle conditions since the single higher performing server has replaced several lower performance servers (Section 2).
3. Many higher performance servers have higher idle power than lower performing servers. A regulation based on the idle power limit and associated allowances will be biased against higher power, higher performing servers and further “tightening” of an idle limit will largely increase the number of high performing servers being excluded from the market (Section 3).
The proposal by DG GROW to define an energy efficient server that’s solely based on its idle power measurement will exclude many high performance servers, as measured by workload delivered per unit of energy consumed, from the EU market. Excluding higher performing servers will mean deployment of lower performing servers in the data centers which will result in net increase of data center energy use, the opposite of the desired outcome of the ErP Lot 9 directive. Several use studies have been cited to justify the use of idle power as an efficiency indicator, citing the low utilization of many servers, particularly in offices and remote “data closets”. Focusing on the individual idle power of an individual server is not an appropriate way to assess how much energy will be used in the data center.
Servers are sized to manage an expected peak workload, recognizing that there will be periods of time when servers are only partially utilized. With the use of virtualization technologies, workloads can be combined onto a single server to increase their utilization, thereby reducing the time they are at low utilization or idle and consequently reducing the number of servers required to do work. Higher performance servers typically have better utilization characteristics, enabling them to consolidate a greater number of individual applications onto a single server. This situation is applicable to both enterprise servers managed in a data center environment and servers installed in an office environment.
Further, tower servers, which are the server class largely expected to be installed in an office environment, represented 20.4% of the market in 2015 and are projected to represent only 16.3% of the market in 2020. Rack server sales are declining as companies increasingly choose to purchase software services through platform or software as a service offerings from cloud computing companies. It is not logical to establish server requirements based on perceived concerns –i.e. servers sitting idle in office environments – when (1) total energy use is dictated by the number and size of the machines deployed to manage the workloads and (2) the servers installed in office environments is trending downward and will represent 16.3% of the market in 2020, around the time the regulation is expected to become effective. Contrary to the comments made by ECOS that “many servers still spend most of their time in an idle state” and “particularly in small and medium server rooms and colocation data centres”, the data collected by DIGITALEUROPE confirms that the market in the scenario described by ECOS represents a minority of the high volume servers sold into the EU market. Colocation and enterprise data centres are complex and support a broad scope of applications and services and a generic statement as described would be inappropriate and not representative.
For these reasons, DIGITALEUROPE continues to recommend that server efficiency should be assessed using the DIGITALEUROPE Server Efficiency Metric (the SERT™ metric or weighted geomean efficiency metric). DIGITALEUROPE and the Green Grid (TGG) have collected a large dataset of SERT™ results and have shown that the metric is effective in measuring server efficiency. DIGITALEUROPE believes that sufficient data has been collected to establish an effective SERT metric threshold for server energy efficiency to remove the least efficient servers from the market and driving reduced energy consumption in the data center (we estimate an energy reduction of 33-78% in power consumption compared to a business as usual case for worst performing servers, see Section 4).
If DG GROW proceeds with its plan to set a server efficiency threshold based on an idle limit, the information presented here demonstrates how higher performance servers would be unfairly and unjustifiably excluded with the potential net impact of higher energy consumption by data centers. Although there is no approach based on idle power limits that would entirely avoid this issue, DIGITALEUROPE has set out below several important changes that need to be made to the idle allowances to lessen the impact:
1. There needs to be an idle allowance for system performance. A low end processor with minimal cores and a lower frequency and a less energy intense supporting infrastructure – system chips, circuitry etc. – generates a lower idle power than a high end processor with maximum cores and frequency and its supporting infrastructure. The higher idle power of approximately 20 watts will deliver performance increases of 200% or more while reducing data center energy use because few higher performing systems will need to be deployed to deliver a given workload. TGG has developed a performance based idle power allowance for each server category identified in the Lot 9 regulation, which when combined with other allowances, eliminates more of the lower efficiency servers as measured by the SERT metric, than the Draft Lot 9 proposal. Incorporating the system performance adder enables a reduction of the base idle allowance to 15 W for a one socket server and 30 W for a two socket server, reduced from the 37 and 85 watts in the Draft Lot 9 document (Section 4).
2. The memory idle allowance can be reduced to 0.175 W/GB. Unlike the ECOS evaluation on one data point from a manufacturer’s on-line product power calculator, TGG has evaluated over 25 data points from two technologies (DDR3 and DDR4), 3 process technology nodes and three manufacturers to justify the lower adder. A summary of the data is provided below (Section 5).
3. The allowances for the storage devices need to be broken into 6 categories by drive type, drive speed, and connection/controller type. TGG has evaluated 284 measured data points to identify the categories and the recommended idle limits. A summary of the data is provided below (Section 6).
4. Idle power allowances for I/O devices need to be added for 25 to 50, 50 to 100, 100 to 200, and greater than 200 Gb/s devices and the allowance for devices with 10 to 25 Gb/s needs to be increased to capture new technologies being released into the market over the Tier 1 period (Section 7).
These changes will better represent the idle power characteristics of the server systems and their associated components than the Lot 9 proposal. They are also based on an extensive set of measured or publically available data for components and systems. The suggestion by ECOS that “there is no contradiction between low-power idle and high-performance” is incorrect. Following that, the statement “the proposed Ecodesign framework is performance-based, with categories and adders that scale power allowance with performance” is also incorrect and it is exactly for this reason that DIGITALEUROPE proposes these changes. DIGITALEUROPE notes that even with these changes, the revised idle power limit will still result in the exclusion of some high performance servers from the EU market even though these servers would yield a lower energy use than some systems that would pass due to lower idle power (see Figs 11 and 12 in Section 3 below).For this reason, DIGITALEUROPE maintains that idle power limits are not an effective or justified means of assessing the energy efficiency of servers.