I, too, read the article incorrectly the first time round. What it's saying is that for an EPYC CPU, performance under systemd is _better_. For a Xeon CPU, performance under systemd is _worse_.

So, the EPYC's architectural differences won't explain why an EPYC CPU runs slower under systemd, because it's the opposite.

Likely bad synchronisation primitive is used, such as sleep instead of optimistic spinning or even locking. Probably not atomic variables where they can be used instead of mutexes.

Alternatively croup confinement causes threads to be migrated at random and some flat overhead.