Definition
Percentage of per-workload latency or throughput degradation events attributable to co-located tenant workload interference rather than to the workload's own demand or the node's aggregate capacity limits.
The Measurement Gap
Measurement Gap and Decision
Decision Supported: Are per-tenant SLA violations being caused by co-location decisions, and if so, which workload pairings are incompatible?
Measurement Gap:
When a tenant workload experiences a latency spike or throughput drop on a shared edge node, the standard troubleshooting path examines two causes: the workload's own demand (did it spike?) and the node's aggregate capacity (is the node overloaded?). If neither explains the degradation, the investigation stalls. The missing third explanation (interference from a co-located tenant workload consuming shared resources (CPU cache, memory bandwidth, network I/O)) is almost never investigated because the monitoring data to attribute degradation to a specific neighbor does not exist. Per-workload resource consumption at sub-second granularity, correlated with per-workload quality metrics, is required to make this attribution. Most edge deployments monitor at the node level, not at the per-workload level, making neighbor interference invisible. NNII formalizes this attribution by correlating degradation events with co-located workload bursts, requiring that the degraded workload's own demand is within normal bounds while a neighbor's resource consumption is anomalous.
Citations:
- ACM Computing Surveys, "Performance Interference in Cloud Computing: A Survey," 2023. Catalogs interference mechanisms but does not propose a standard computable metric for operational use.
- ETSI GS NFV-TST 008, "NFV Performance Test Report," 2024: defines performance testing for virtualized functions without measuring cross-tenant interference in production environments.
Formula
Plain Language: For each tenant workload experiencing a quality degradation event (latency spike, throughput drop), determine whether the degradation correlates with a burst from a co-located workload. If the degradation occurs simultaneously with a co-located burst AND the degraded workload's own demand is within normal bounds, attribute the degradation to noisy-neighbor interference.
Notation:
NNII = Interference_attributed_degradation_events / Total_degradation_events × 100
Variable Definitions:
- Interference_attributed_degradation_events = quality degradation events for a workload that correlate with a co-located workload burst AND the degraded workload's own demand is within historical norms (count)
- Total_degradation_events = all quality degradation events across all tenant workloads (count)
- Attribution requires three conditions: (1) temporal correlation with co-located burst, (2) degraded workload demand within normal range, (3) co-located burst consuming shared resources (CPU cache, memory bandwidth, network I/O)
Units:
Percentage. 0% = all degradation is self-caused or node-capacity-caused. High NNII = co-location is a significant source of SLA violations.
Sign Convention:
Lower is better. 0% means co-location interference is not a measurable problem.
Edge Cases:
- Multiple co-located workloads burst simultaneously: attribute to the workload(s) consuming the most shared resources, not all co-located tenants.
- Infrastructure-caused degradation (node hardware issue): exclude. This is not a neighbor interference event.
- Attribution requires sub-second correlation. Coarse monitoring (1-minute intervals) will miss brief interference events.
- Minimum sample: 50 degradation events per measurement period.
Proposed formula, pending validation pilots.