Sliding windows separate receiver space from path capacity

The sender tracks a first unacknowledged position and a next position to send. Their difference, after correctly accounting for sequence wrap and recovery state, describes the outstanding stream span in a simple loss-free trace. A receiver advertisement allows a range of acceptable positions beginning at its cumulative acknowledgment. As acknowledgment advances, the left edge moves forward and older bytes stop occupying outstanding credit. Do not add every observed window advertisement together: a new advertisement describes a current window, not a new grant unrelated to all earlier grants.

For a simple steady connection outside recovery, the smaller of the congestion window and receiver window bounds outstanding data. Subtract already outstanding bytes to estimate how much new payload can be sent. This calculation also assumes the application has queued data and that pacing, local buffers, and implementation rules permit immediate transmission. A window measured in bytes is not itself a bitrate. Divide an appropriate usable flight limit by feedback time only as an approximate throughput ceiling, and then check whether application behavior or bottleneck capacity is tighter.

Receive credit reflects data the endpoint is prepared to accept while bytes await application consumption. If the application stops reading, queued data can fill available receive space and drive the advertised window to zero. Increasing the network bandwidth cannot make that application consume its buffer. Investigate blocked readers, backpressure, memory policies, and application execution before changing congestion control. A larger receive buffer might absorb a short pause, but it can merely postpone a persistent mismatch while retaining more memory and increasing the age of queued work.

The TCP header's basic window field has sixteen bits. Window scaling negotiated during connection establishment permits larger advertised windows by applying a directional shift to later window values. The SYN's own window field is not scaled. Each direction can negotiate a different shift, so a capture must retain handshake context to decode the field correctly. Configuring large host buffers does not guarantee a correspondingly large effective advertisement; the negotiated scale, current receiver state, and implementation policy all matter. Record the effective byte count before comparing it with a path's bandwidth-delay product.

A zero advertisement means the receiver currently offers no new ordinary receive space. TCP's zero-window probing prevents a lost window-opening update from causing an indefinite mutual wait. A responsive peer can keep advertising zero for an extended period; that behavior differs from an unreachable peer or an application-level timeout. Applications still need their own resource and deadline policies. Holding a connection forever can consume descriptors and request state even when the transport is doing its job. Log the time spent receiver-limited separately from loss recovery and application inactivity.

In the queue lab, interpret rates as normalized byte units per second and service as reader consumption. Select Reset, explicitly set both λ=3 and μ=2, predict growth from four units at t=0, then Play or Step. For a separate drain comparison, Reset again, re-set both λ=1 and μ=2, and predict zero backlog at t=4 seconds before playback. Parameter edits retain elapsed time and recompute from the initial four; they do not continue a changed-rate history. The model lacks finite receive credit and sender feedback. Its separate M/M/1 residence estimate is not an exact TCP drain time.