According to current surveys, 58% of datacenter operators do not enable or exploit the Energy Savings which can be readily derived from the Dynamic Voltage and Frequency Scaling features which are implemented in all modern servers powered by AMD and Intel processors.

---

What’s a Watt?

The Watt is a unit of power equal to 1 joule per second. Given that, the product of (1 Watt * 60 seconds * 60 minutes) delivers a measure of energy termed a Watt-hour, abbreviated "Wh" and equivalent to 3,600 joules.

For all practical purposes of utility billing, 1 Wh is an extremely small quantity. Consequently, utility billing will typically reflect electrical consumption in units of kiloWatt hours (1,000 Wh) which is abbreviated "kWh", or megaWatt hours (1,000,000 Wh) which is abbreviated "MWh".

---

DVFS Overview

Dynamic Voltage and Frequency Scaling (DVFS), also known simply as Dynamic Voltage Scaling (DVS), is a technique for varying the voltage and clock frequency of microprocessors and their cores in order to produce a reduction in energy consumption.

Typically, the frequency and voltage are repeatedly altered in realtime by a software-implemented algorithm designed to match processor speed to computation load, ideally producing energy savings with little or no impact on application performance.

DVFS may also be employed to achieve other energy-saving objectives such as 'throttling', or confining energy consumption and thermal output to targeted limits.

The Roles of Hardware and Software

Silicon designers have provided our industry with hardware features that can be exploited to implement DVS algorithms; in other words, DVS energy savings is not implemented by hardware, but is, rather, enabled by hardware features.

DVS algorithms and their related functionality are implemented in specialized system software generally termed 'power governors'. For example, most versions of Unix, Linux, and Solaris provide the requisite kernel support software ('CPU drivers') and power governors to achieve varying levels of energy savings through DVS.

However, it is left up to the IT consumer to enable and exploit DVS capabilities when provisioning servers. This is a step which remains poorly understood, and in 2006, it is reported that 58% of datacenter operators did not evaluate and appropriately exploit this capability.

Top of Page

Obtaining Energy Savings

In order to visualize the region in which DVS algorithms produce energy savings, an example of the results of a TPC-C performance benchmark from a recent Trilliant Group Server Energy Assessment is presented in Figure 1. The target servers were 4-way single-core servers based on 3.6 GHz Intel Xeon processors with 16GB of DDR-2 memory. As can be seen, the TPC-C benchmark provides sufficient threads to present a significant computational load to each of the four processors.

Figure 1 - TPC-C Computational Load

The results in Figure 1 are plotted at 60-second intervals over the course of the 21-minute test period. By continually moving the processor frequency to deliver only the MIPS required to execute the benchmark, the power governor 'mined' energy savings from the region above the sawtooth.

The atypically high average computational load (82.5%) provided only a small band for energy reduction, yet signifcant energy savings were nevertheless achieved. Performance impact was nil, and the net result of this test was that average power demand was reduced from 634 to 564 Watts; and the energy consumption for the test period was reduced from 222 Wh to 199 Wh -- a processor energy reduction of over 10%. Such savings might conservatively translate into over $100 in energy savings per year for each server.

Please see our Reference Library for more information on DVS from AMD ("Cool'n'Quiet Technology") and Intel (Enhanced Intel SpeedStep, or "EIST").