As reported in the March 26th issue of GreenerComputing News (you do subscribe, right?), the Standard Performance Evaluation Corp. (SPEC) is in the process of formulating a power-consumption benchmark for workstations. SPEC is a vendor-neutral, non-profit organization that designs benchmarks for the computer industry. It also hosts a website, www.spec.org, that presents benchmark results for various platforms. Those results are provided by vendors of hardware and software systems who certify that they ran the benchmarks in accordance with SPEC guidelines.

While SPEC cannot and does not vouch for those results, the industry puts real effort into maintaining the integrity of posted results. And because those results often figure prominently in marketing campaigns and reports by market analysts, competitors track each others' numbers carefully, which provides an additional incentive for fair, accurate reporting.

The SPEC benchmarks are almost always suites of tests that exercise the key features of a hardware or software component. A good example is the SPEC ViewPerf benchmark that I discussed in my previous column. It measures the performance of graphics subsystems on PCs and workstations by running through a series of visual tests. As I mention in that column, ViewPerf can be downloaded from SPEC and run to generate a series of results that enable comparison between two graphics cards.

Returning to power consumption, the issue faced in designing benchmarks is how to create a realistic suite that generates numbers that accurately reflect the quantity being measured. For example, how to reflect power consumption of the "typical" use model for a PC, workstation, or server?

For servers, SPEC formulated an interesting response to this question when it designed the SPECpower_ssj2008 benchmark suite. Because there are so many different kinds of servers, a universal profile would have been very difficult to develop. (Consider, for example, the different usage levels for a Web server than, say, an authentication server -- the latter being used exclusively to check user logons and passwords.) SPEC chose a Java application running on the server.

The software generates results for the server under workloads starting from 0 percent load (no activity) rising to 100 percent load in 10 percent increments. The benchmark tracks the number of operations performed at each workload level and the power consumed. The average of test operations per watt consumed at each power level is the ultimate benchmark figure. Figure 1 shows this data for a Dell PowerEdge 1950 III server (with dual Intel Xeon E5440 processors).

Figure 1. How the final result is computed for SPEpower_ssj2008 using ratings for every 10 percent of workload. (Courtesy: SPEC)

A key point in this calculation is that by taking the average of the numbers at 11 levels of work load, SPEC modeled servers as spending equal time at each level. This activity profile probably matches no single server, but is a reasonable model for measuring all servers.

Now, if we look at workstations, the model to use to get a representative number is considerably more elusive. Workstations are generally associated with two kinds of special activities: number crunching and high-end graphics. In addition, workstations generally have more disk drives than comparable PCs. The difficulty in factoring all these elements together can be seen in other benchmarks that try to capture workstation consumption levels. For example, the recently released Energy Star v. 4 specification of Requirement For Computers (http://www.energystar.gov/ia/partners/product_specs/program_reqs/Compute...) has straightforward power-consumption measures for PCs and even servers, but it relies on multivariable formulas when it comes to workstations. To qualify for the EnergyStar certification, a workstation's average power draw must be less than or equal to:

0.35 (Pmax + (HDD x 5) watts

where Pmax is the maximum power the workstation can consume and HDD is the number of hard disk drives. Coincidentally enough, the Pmax figure can be generated (according to the EPA) by running the SPEC ViewPerf benchmark mentioned earlier plus the Linpack benchmark (found at http://www.netlib.org/linpack/). The problem with Linpack is that it's a Fortran benchmark and if you can't compile Fortran, you can't run it. However, a Java version can be found here: http://www.netlib.org/benchmark/linpackjava And Java compilers are available from Sun for most platforms today.

Taken together, the formula and the two benchmarks enable you to measure the power consumption of your workstations (and, if you want, your high-end PCs). The only missing element is the meter to measure the watts consumed. As mentioned in previous columns, I recommend the Kill-A-Watt Electricity Usage Meter from P3 International, which is inexpensive (around $25) and easy to use.

Now, you can establish your own set of benchmarks for power consumption at your site without waiting for SPEC to complete its own benchmark suite.