The idea is to exclude as many sources as possible that could lead to issues with contention and cache misses. After some experimentation, we think that hyperthreading is simply not working very well with our kind of workload. In the future we might simply test on other hardware.
