Notes:

• To run prefetchers' code in the simulation framework, please download the simulator first, then copy the .cpp and .h files to sample_prefetcher.cpp and sample_prefetcher.h in the PREF_KIT/src/prefetch/ directory, and follow the instructions on the webpage: Simulation framework download and directions
• Results were obtained using simulations of the SPEC CPU 2006 benchmarks only, and do not include any commercial benchmarks. Please take that into account when evaluating results and ranking.
• To generate the results, we obtained traces of SPEC CPU 2006 benchmarks that were compiled with the icc compiler. We used the first reference input, fast-forwarded for 40 billion instructions, and obtained a trace for 100 million instructions.
• A few contestants submitted codes that had bugs where programs go to infinite loops, which prevented some simulations from running to completion. Rather than disqualifying the submission, the organizing committee chose to use a speedup of 0.667 for configuration 1 and a speedup of 0.5 for configurations 2 and 3 in place of the incomplete simulations. However, this did not have an impact on the performance ranking. The performance champion did not have any incomplete simulations.
• We computed the speedup over no prefetching for each simulation run across all benchmarks. Each contestant got three scores, one for each configuration, representing the geometric mean of the speedup (over no prefetching) across all benchmarks. The final performance score was the sum of the scores for the three configurations.
• As a reference, the “no prefetching” technique has a score of 1.0 for each configuration, for an overall performance score of 3.0. See all no prefetching results here
• As another reference point, we also distributed a rudimentary PC-based stride prefetcher with the simulator. That simple prefetcher had scores of 1.14, 1.07 and 1.06 for the three configurations, for an overall performance score of 3.27. See all stride prefetching results here