You are solving a design failure on a massive Hadoop cluster network that has an application with TCP incast behavior (also known as TCP Throughput collapse) affecting its many-to-one communications w

Availability measures whether the network is reachable and does not capture the micro-burst buffer saturation that drives incast collapse.

Bandwidth utilization reflects average throughput but not the instantaneous burst that fills switch buffers during many-to-one incast events.

Jitter measures packet delay variation and is relevant to latency-sensitive real-time traffic, not to buffer-overflow-induced TCP throughput collapse.

TCP incast occurs when many servers respond concurrently to a single requestor, filling the ingress buffer of the last-hop switch and causing tail-drop packet loss. This triggers TCP retransmit timeouts, collapsing throughput. Measuring buffer utilization at the aggregation switch identifies whether buffers are saturating and guides QoS or buffer tuning to mitigate the issue.