TIL: using hpet (instead of tsc) as #Linux clocksource apparently can cause massive slowdowns in some workloads. E.g. when reading data from a socket using select() and read(), with hpet clock the select() calls alone can reduce throughput down to 25%. #Perf profiling shows that the kernel spends lots of time in read_hpet().
Guess now I cannot postpone debugging the clock drift problems that occur with tsc clocksource on this one system :-(