We've hit a performance problem with multi-socket setups (NUMA) - basically one core talking to the other's core memory (different sockets), hence slowdowns. And it's amplified by the use of atomics (shared_ptr, etc.). What we did was simply to disable one core (it was two CPU sockets). I guess for servers, executing many (mostly) single threaded jobs (a.k.a "borg", "kubernetes", "docker") is okay, but when you have app using all threads (TBB-style) then it becomes bottlenecked.

SysInternal's https://docs.microsoft.com/en-us/sysinternals/downloads/core... actually captures this detail (probably there are better apps, but it was nice to inspect machines quickly). The tool measure some relative slowdown.