SPECjbb2005 took somewhat more analysis than our other benchmarks, due to its construction. In particular, the benchmark varies the number of warehouses, each one corresponding to a thread. A script steps through different warehouse levels, from 1 to twice the number of hardware threads supported (usually). Figure 4 below shows the power consumption for both Dempsey and Woodcrest. Note that Dempsey, by virtue of multithreading, requires twice as long to run the benchmark, and this is in no way an indicator of performance.
Figure 5 – SPECjbb2005 Power Consumption
For SPECjbb2005, the initial ‘warm up’ warehouses (1-3 for Woodcrest, 1-7 for Dempsey) run for 30 seconds each. The ‘measured’ warehouses (4-8 and 8-16 respectively) execute for 2 minutes, and will dominate the run time of the benchmark. Looking carefully at the execution profile, it is easy to see where the benchmark is executing and the regular intervals where the script is exiting from a run and setting up for the next set of warehouses. To provide performance/watt, we only averaged over the ‘measured’ warehouses, so the warm up periods were ignored, as they don’t contribute to the score.
Chart 5 – SPECjbb2005 Summary
For SPECjbb2005, the change in power draw is in the same neighborhood as previous benchmarks, at 132W. However, the overall performance in SPECjbb2005 did not increase substantially, only around 20%. This brings the overall improvement in performance/watt to 1.85x, lower than most of the other benchmarks.