• Sources: LWN, slides, HN discussion
  • Summary: The talk and slides list specific behaviours that cost PostgreSQL throughput: futex hash lookups taking 25 percent of CPU in some cases, a throughput plateau that disappears when cpuidle is disabled because the load is not enough to keep CPUs in a high-power state, no transparent huge pages for program text now that READ_ONLY_THP_FOR_FS is being removed, vm.overcommit_memory=2 not working inside a control group, fallocate reservations causing fragmentation and journal writes, 1GB huge pages halving performance through folio-head refcount contention, and io_uring registered buffers capped at 1GB per operation.
  • Why it matters: Most of these are properties of process-based servers on Linux rather than of PostgreSQL, so the same items apply to any such workload.