Transparent huge pages with defrag set to madvise

A service sets THP's defrag mode to 'madvise' specifically to avoid the well-known page-fault-path stall, where a thread that faults in new memory can itself block on synchronous compaction while the kernel assembles a contiguous 2MB region. Even with that mitigated, the team still occasionally sees a few-hundred-microsecond latency blip on a completely unrelated, already-warmed-up hot-loop thread of the same process, a thread that isn't touching any new memory and didn't trigger any fault itself. What's the most likely cause?

37 of 53 in Operating Systems