In the router. Compute a better default for the number of processors based on multiple observations: given enough streams to occupy all processors, performance is degraded if there are as many processors as cores. Best performance is obtained by setting aside two cores for other tasks. Making this the default simplifies the benchmarking process by not having to configure an override for each type of target machine being tested.
In the benchmark. Run the test with many streams. It reflects reality and allows a very parallel router to perform at its best without distorting the performance of a less-parallel one. Make sure that the number of streams is likely to be a multiple of the number of cores so the cores are loaded evenly in spite of there being typically fewer streams than in real life. (TBD, learn the router's internal processors count and just use that).
In the router. Compute a better default for the number of processors based on multiple observations: given enough streams to occupy all processors, performance is degraded if there are as many processors as cores. Best performance is obtained by setting aside two cores for other tasks. Making this the default simplifies the benchmarking process by not having to configure an override for each type of target machine being tested.
In the benchmark. Run the test with many streams. It reflects reality and allows a very parallel router to perform at its best without distorting the performance of a less-parallel one. Make sure that the number of streams is likely to be a multiple of the number of cores so the cores are loaded evenly in spite of there being typically fewer streams than in real life. (TBD, learn the router's internal processors count and just use that).
This change is