'

Check out the SPEC: New benchmark hones in on server power consumption vs. performance

The Standard Performance Evaluation Corp. (aka SPEC) has whipped up a new benchmark that measures the power consumption of a given server relative to the job it’s doing.

The Standard Performance Evaluation Corp. (aka SPEC) has whipped up a new benchmark that measures the power consumption of a given server relative to the job it’s doing.

Called the SPECpower_ssj2008, the benchmark toolset costs $1,600 if you buy it from the consortium (although non-profits and educational institutions can get it at a discount).

The benchmark works by reporting power consumption for different performance loads from idle to 100 percent (in 10 percent increments). It is computed by adding together the transaction throughputs for each level and then dividing them by the power consumed for the segment. SPEC calls this measurement the “overall ssj_ops/watt.” The current benchmark tests for a variety of multithreaded workloads associated with running server-side Java business applications. Different types of workloads will be added.

So, for example, you can figure out whether running an existing server at full capacity is really the best thing for either power or performance or rather investing in a virtualized architecture might be better for the environment and your company's bank account.

Klaus-Dieter Lange, chair of the SPECpower committee, says about a dozen vendors have released results for their various servers, which you can find here. Several more results are expected by the end of January.

The information likely will be used by Energy Star, which is considering a new weighting for server that includes updated power consumption information, according to Lange. The consortium also is sharing its work with the Environmental Protection Agency and with the Consortium for Energy Efficiency. The latter organization is reviewing the benchmark with an eye to encouraging customers to look at industry-endorsed technology power ratings more closely.