A strategy thread is pinned to an isolated core and still shows a 250 microsecond stall roughly once a minute. It performs no allocation, no logging and no system calls in the loop. Give four distinct causes that survive core isolation and how you would confirm each.

A strategy thread is pinned to an isolated core and still shows a 250 microsecond stall roughly once a minute. It performs no allocation, no logging and no system calls in the loop. Give four distinct causes that survive core isolation and how you would confirm each.

Approach: Enumerate the things that can steal a core without the scheduler running another task on it, then match each to a counter or trace that proves it.

System management interrupts, TLB shootdown inter-processor interrupts, page faults on cold or swapped pages, and cache or memory bandwidth interference from other cores all survive isolcpus. A system management interrupt is taken in firmware below the operating system, is invisible to every kernel counter, and can last hundreds of microseconds; confirm it by reading the SMI counter in MSR 0x34 before and after a stall window. A TLB shootdown happens when another thread in the same address space unmaps memory, which sends an inter-processor interrupt to every core using that mapping including yours; confirm it in the per-core interrupt counts for the TLB and call function categories. A minor page fault on first touch or a major fault on a page reclaimed under memory pressure costs microseconds to milliseconds; confirm with the process fault counters and remove it with mlockall and pre-touching. Interference is a noisy neighbour on another core evicting your lines from the shared L3 or saturating the memory controller; confirm with the per-core miss counters and fix it with cache allocation technology to partition L3. Tail latency work is this list, one cause at a time, since none of them show up in a mean.

Follow-up: You cannot disable SMIs on this platform. What design change bounds their damage to a single strategy rather than the whole desk?

Key concepts: system management interrupt, page fault, tlb shootdown, tail latency.