Transparent huge pages and latency spikes

A latency-sensitive service enables Transparent Huge Pages (THP) in 'always' mode to reduce TLB misses on its multi-gigabyte heap. In production it occasionally sees multi-millisecond stalls that don't correlate with GC pauses, I/O, or lock contention. What is the most likely cause, and how does switching to explicit hugetlbfs pages avoid it?

20 of 53 in Operating Systems