ACG-RR: Confidence-Gated Failure Detection and Grouped Route Repair for Data Center Fabrics

Failure detection that is always aggressive amplifies false route withdrawals during short congestion bursts, whereas conservative detection increases the outage preceding route repair. This paper presents ACG-RR, an event-driven, confidence-gated monitoring and grouped route-repair policy for leaf–spine data center fabrics. Following the simulate–analyze–validate workflow established for data center fabric convergence studies, we implement a synthetic failure-event simulator, compare seven precisely defined policies, and gate every reported result on a 15% agreement check between event-integrated loss estimates and independently drawn packet-time samples. The evaluation comprises 128 paired topology/failure/seed scenarios (896 policy runs), including single-link, spine, correlated, and two-wave failures. ACG-RR yields 46.1 ms mean scenario-level 95th-percentile blackout versus 116.4 ms for 25-ms fixed probing and per-flow repair (a 60.4% reduction), while using 35.9 probes per link per second. An aggressive 8-ms detector with grouped repair is faster (41.9 ms) but generates substantially more false alerts and modeled traffic disruption. All measurements arise from fully disclosed simulated workloads, not production packets or a hardware testbed. These results quantify a controllable speed–stability trade-off and identify when repair grouping dominates timer tuning.