Incast and the budget before feedback

Draw the fan-in boundary: several ingress paths converge on one output link or receiver. During the overlap, aggregate arrivals can exceed that boundary's service rate even when every upstream link is below its own capacity. A finite buffer absorbs the difference temporarily. Once packets are dropped or paused, the source traffic changes, so a constant-rate estimate describes an initial interval rather than a complete transport execution. State the overlap duration and starting occupancy explicitly.

Average utilization loses temporal structure. A burst lasting one hundred microseconds may be nearly invisible in a one-second mean while producing a queue that delays every following packet. Flow completion can depend on the slowest worker, making a single retransmission timeout costly to the whole request. Distinguish application fan-in, packet arrival synchronization, and receiver processing: slow receive queues or memory paths can create symptoms similar to a congested switch output.

A congestion signal must travel back and cause senders to reduce future transmission. Meanwhile already transmitted bytes continue arriving. Headroom must account for that reaction interval and the actual number of converging senders. A threshold that is safe for one rack's geometry may be unsafe after link speeds, cable distance, or scheduling change. Increasing buffer space can absorb a larger burst but also allows longer waiting when congestion persists; it does not create additional egress bandwidth.

Application-aware fan-in scheduling can limit concurrent responses, while pacing spreads them across time. Both trade scheduling delay and coordination against peak queue pressure. Smaller chunks may improve responsiveness but increase overhead. Compare complete request latency, tail behavior, and throughput under the real fan-in distribution rather than optimizing only a bulk-transfer test. Include background traffic because a buffer partly occupied by another workload has less room for the same synchronized replies.

Measure high-resolution occupancy or device-supported watermarks alongside drops and per-priority traffic, while recording fan-in and message size. A zero drop count at the switch does not prove no congestion: link-level pauses, endpoint backpressure, or sampling gaps may hide it. A full buffer counter also needs a definition; shared-pool accounting and reserved headroom can differ from an advertised total memory size. Locate the actual bottleneck instead of treating all buffering as interchangeable.

For an abstract comparison, Reset the queue lab, set both λ=4 and μ=2, and predict backlog six after one time unit; two Steps reach that time from the initial four jobs. Reset and re-set λ=2, μ=4 before predicting backlog two at the same time. The lab uses fluid jobs, not packet bytes, has no finite buffer or feedback, and cannot report losses. It illustrates only how excess arrival rate integrates into backlog under fixed rates.