Problem

finalizeOperationStats accounts for 0.53% flat CPU (0.68% cumulative) on the 95%-read path (CurOp::completeAndLogOperation → TicketingSystem::finalizeOperationStats → ExecutionAdmissionContext::finalizeStats). For every completed operation, the function unconditionally executes ~40 fetchAndAddRelaxed operations on shared global AggregatedExecutionStats counters. On the 95% read path, the vast majority of these atomics add zero — write-path counters (writeShort, writeLong, writeDelinquency) are always zero for reads, and long-running counters (longRunning, readLong) are always zero for short-running reads. On ARM64 (Graviton), each atomic RMW instruction generates exclusive-store traffic on the cache line regardless of the value added, causing unnecessary cache-line bouncing across all 128 worker threads sharing the _operationStats structure.

Solution

Guard each accumulation block in finalizeOperationStats with a cheap per-call check before touching the shared global counters. Each guard tests a single loadRelaxed field from the per-operation snapshot that serves as a provably complete proxy for "all fields in this group are zero" — for example, totalOpsFinished == 0 for the finalized-stats group and totalAdmissions == 0 for the per-acquisition groups. When all source fields are zero, the entire AggregatedExecutionStats::operator+= call (and any nested incrementDelinquencyStats call) is skipped entirely, eliminating the unnecessary cache-line writes without changing observable behavior.