Performance Per Watt: Dempsey and Woodcrest

Pages: 1 2 3 4 5 6 7

XML Test 1.1

XML Test 1.1 is a moderately complicated benchmark to handle, primarily because it is composed of 9 subtests, which use two different methods (Document Object Model and Simple API for XML) for serving XML documents using the 32 bit version of BEA’s JRockit. Each subtest requires around 500 seconds, and fully loads the processors with the appropriate number of threads.

Figure 4 – XML Test 1.1 Power Consumption

One rather surprising element is the oscillating power consumption. For most of the other benchmarks, the power consumption appeared to be nearly constant, while here it varies substantially and randomly in some cases. The power consumption for the first six tests (using DOM), on Woodcrest is 240-250W, while Dempsey jumps around between 335 and 370W. For the latter three SAX tests, the power consumption for Woodcrest appears to converge to ~260W, while Dempsey settles on a trimodal sort of distribution, clustering around 325, 360 and 395W. The most probable explanation for these variations is that XML Test 1.1 includes garbage collection, which may change execution characteristics for short periods of activity. It also seems very likely that the multithreading in Dempsey exacerbates this variability, perhaps causing some interaction between extra garbage collection threads, computational threads and the Windows scheduler.

Chart 4 – XML Test 1.1 Summary

Performance for XML Test was calculated by averaging the results of all 9 tests. According to this straightforward approach the performance difference was 32%. Out of all the tests, XML Test showed the smallest difference in power consumption, around 94W. Woodcrest showed a substantial improvement in performance/watt, a factor of 1.81x, but the smallest of all the benchmarks.

Pages: « Prev   1 2 3 4 5 6 7   Next »

Discuss (68 comments)