Arista EOS insights

When Connectivity Monitor detects abnormal latency to configured hosts (exceeding historical bounds by configured anomaly score), it signals potential upstream network path issues, cabling problems, or remote endpoint degradation. Default threshold is 100 standard deviations; breaches generate syslog events that should trigger investigation of physical layer and routing paths.

Strata chipsets (7010/7050X/7060X/7250X/7260X/7300X series) generate capacity warnings when TCAM slices exceed 90% utilization. Unchecked growth generates high-volume syslog events and risks feature failures when tables reach 100%. Thresholds are configurable per-table and per-slice to reduce noise while maintaining visibility into resource exhaustion.

PTP monitoring thresholds for offset-from-master (max 1 second), mean-path-delay (max 1 second), and skew (0-1000% frequency difference) generate syslog alerts when time synchronization degrades. For applications requiring microsecond-level timing accuracy, threshold breaches require immediate investigation of master clock health, network path asymmetry, and physical layer issues.

Abnormal interface RX/TX errors or discards (detected via anomaly scoring) in AI cluster fabrics directly impact distributed training job completion time (JCT). Even minor physical layer issues cause step-time jitter and GPU idle time, as AI frameworks cannot tolerate packet loss like traditional TCP workloads. CloudVision CV UNO correlates interface health with AI job metrics for root cause analysis.

CloudVision CV UNO's AI Jobs Dashboard reveals when distributed training job completion time (JCT) degrades due to network issues. Correlation with LANZ (sub-millisecond congestion tracking) and Traffic Overview Dashboard identifies specific congestion points, flow imbalances, or RoCE misconfigurations causing GPU idle time and reduced training efficiency.

Health Monitor Alerts Indicating Infrastructure Faults

System health status degraded due to detected faults in storage, network, or cluster components. Health monitors track critical conditions like disk failures, switch connectivity issues, or HA interconnect problems.