In December 2007 SPEC, an industry standards and benchmarking organisation, released the first standard for measuring and reporting power consumption of servers. The SPEC standard is a benchmarking tool - not a definitive answer to measuring power usage - however it does provide data appropriate for "a comparison point between offerings in a specific environment".
SPECpower specifically measures the watts (power) of 1 type of load (a java application running pre-defined scripts) at Active Idle, 10, 20, 30, 40, 50, 60, 70, 80, 90 and 100% load. The load is generated by scripts provided by SPEC (at a cost) and the result is a set of transaction throughputs, called ssj_ops, at each load segment e.g.
| Load |
ssj_ops |
Average Power (W) |
| 100% | 88,566 | 321 |
| 90% | 76,121 | 315 |
| 80% | 71,222 | 301 |
(etc)
At the same time each load segment is measured, so is the resulting Watts used (using an external power analyser). The sum of ssj_ops and the sum of the power measurements are then totalled and ∑ssj_ops / ∑power = SPECpower_ssj2008 measurement i.e. a single figure able to be used for benchmarking.
This all sounds a bit complex right? Well it kind of is...
A standard for measuring server power consumption has been hard to achieve to date, due to the variances in measurement. There are a number of organisations and universities working on standards for power consumption measurement in the technology realm however it is generally far more complex than it might initially seem. A few considerations for measuring power include:
- What applications are installed and used, and what is the configuration of a server - it is theoretically possible to have two PCs with exactly the same hardware and the same applications installed, but with different configurations causing it to use different amounts of energy (most notably in direct power influenced configurations such as power settings).
- The design of the server - a server with a streamlined design may require less cooling and thus uses less energy. SPEC note that even servers that have the same 'number' of hardware components may differ in power use due to the hardware design - e.g. two servers with two hard drives each of the same specification, but because one of the servers is 'designed' to support more hard drives (even though the drive slots are empty) it may draw more power.
- The components - a server with 1 processor versus a server with 4 processors will require different amounts of energy, and the same goes for the speed of those processors. Additionally the amount of memory installed and type, the number of hard disks installed and their type etc affect power use.
- The amount of data being processed - a server running an application at 10% processing power versus a server running an application at 75% processing power uses a different amount of energy because it is pulling different amounts of power via the hardware e.g. increased CPU = increased power use. This is why the SPEC standard measures at all loads and sums the results to provide an overall 'power' figure.
- Environmental conditions - such as air density, temperature and the method of cooling will also affect power measurement i.e. the lower the temperature the lower the power use.
There will be other power measurements, using different applications for benchmarking in the future however SPECpower_ssj2008 provides a good overall baseline for hardware vendor benchmarking.
References:
- SPEC 2008, SPECpower_ssj2008 Methodology
- SPEC 2008, SPECpower_ssj2008 FAQs
- SPEC 2008, SPECpower_ssj2008 Results
