All good kernel micro-benchmarks will involve a ton of volume, and thus be exercising throughput. All the micro-benchmarks Ingo used are extremely latency-focus

All good kernel micro-benchmarks will involve a ton of volume, and thus be exercising throughput. All the micro-benchmarks Ingo used are extremely latency-focused: a few processes blocking on Pipe I/O or messaging as frequently as possible -- the only subject of the test is how fast the scheduler can toggle between the processes. BFS was dramatically worse across the board in all the benchmarks, except curiously when the # of processes == # of logical processors -- to me that's evidence of extreme naivete in its design.