How do you boost server efficiency? Buy new kit, keep it busy

Increase in efficiency between 2017 and 2019 hardware is 34% for Intel servers, 140% for AMD, claims report

Organizations looking to cut power usage within datacenters should find that newer servers can offer decent energy efficiency improvements, but that efficiency increases with processor utilization, so those servers need to be kept busy.

Datacenters and the networks that interconnect them now account for up to 3 percent of global electricity use and nearly 1 percent of energy-related greenhouse gas emissions, according to figures from the International Energy Agency (IEA).

In addition to this, the recent soaring energy price inflation has caused many organizations to think more seriously about holding down their energy use.

Servers can account for more than half the energy consumption of a modern datacenter, according to a new report from the Uptime Institute, with the remainder used for storage, networking and other facilities such as cooling. This makes server power efficiency a key area for enterprises to focus on when striving to reach sustainability targets, it states, and offers some advice.

The report is based on an analysis of data from the Server Efficiency Rating Tool (SERT) database of the Green Grid, a non-profit industry consortium working to improve datacenter efficiency. The report also only considers x86 servers.

Newness really counts

The first insight is that server efficiency improves with each new generation of hardware. This may seem obvious, but the report has the numbers to demonstrate its point, comparing servers from 2017 with those from 2019 and 2021.

At 50 percent CPU utilization, the increase in efficiency between the 2017 and 2019 hardware is 34 percent for Intel servers and 140 percent for AMD-based servers.

Between the 2019 and 2021 hardware generations, Intel servers saw a 32 percent increase in efficiency, while the figure is 47 percent for AMD.

Comparing Intel against AMD, the Uptime Institute found that the average 2021 AMD server is 74 percent more efficient than the average 2021 Intel server at 50 percent CPU utilization, and 67 percent more efficient at 25 percent CPU utilization.

However, the report cautions that these are average server efficiency figures based on one of The Green Grid SERT tests which uses server-side Java (SSJ) to simulate enterprise online transaction processing applications.

The Uptime Institute also reports that the average efficiency of a server increases by at least 50 percent when CPU utilization is doubled from low levels (20 to 30 percent) to higher levels (40 to 60 percent).

This means that efforts to increase the utilization of existing servers can be a viable alternative to purchasing new servers for some organizations. But combining a utilization increase with a server refresh, efficiency improvements greater than 100 percent can be achieved.

The upshot of that is that more than twice the workload can be processed for the same amount of energy, providing a considerable reduction in energy requirements when done at scale, according to the report.

In contrast, replacing every server with a new one running the same workload will not show anything like the same benefit, resulting in decreased server utilization and much smaller efficiency improvements, in the region of 10 percent.

Another finding of the report is that newer server hardware is most efficient when configured with higher core counts, and this holds for both AMD and Intel systems, although the efficiency improvement is less distinct for Intel servers than for AMD.

Coming in hot

However, CPU utilization must be kept up to benefit from the full efficiency gains of a server with a high core count, and the report says this may be difficult to achieve for some workloads or combinations of workloads.

This will not be news to the hyperscale companies and the big cloud operators, which have been driving the trend towards processors with more cores, such as the upcoming AMD 128-core Epyc processor codenamed Bergamo that has been designed specifically for cloud-native workloads.

The Uptime Institute also found that power management systems can bring efficiency improvements of at least 10 percent, although it notes these are rarely applied in real life because of a real or perceived impact on compute performance.

There is perhaps some justification for this: the report states that power management delivers an idle power reduction of 41 percent, but also reduces server work capacity by 5 to 6 percent, which holds across all utilization levels.

In servers based on AMD Epyc 7262 processors, the largest efficiency savings from using power management are about 15 percent to 19 percent, and are seen when running at 25 to 60 percent utilization, but drop away for CPU utilizations above and below this.

Power management can also increase latency by between 20 and 80 microseconds, the report states. This might sound small but could prove unacceptable for latency critical workloads such as financial trading and may increase the risk of service level agreement (SLA) violations.

In summary, the Uptime Institute finds that deploying newer high-core servers helps to maximize server efficiency, but only if those servers are kept busy. The key to datacenter energy efficiency is improving server utilization, it advises. ®