Network & Telecom

Congestion Control

Q: How does TCP congestion control differ from flow control?

Flow control protects the receiver: the advertised receive window (rwnd) tells the sender how much buffer space the destination host has free, preventing a fast sender from overrunning a slow application. Congestion control protects the network between the endpoints: the congestion window (cwnd) is the sender's own estimate of how much unacknowledged data the path can hold without overflowing intermediate router queues. The sender transmits the minimum of cwnd and rwnd. cwnd is grown and shrunk dynamically from feedback signals such as duplicate ACKs, retransmission timeouts, or ECN marks, whereas rwnd is simply reported by the receiver. On a congested wireless backhaul link the limiting factor is almost always cwnd, not rwnd.

Q: Why does TCP congestion control behave poorly over wireless links?

Classic loss-based algorithms such as Reno and CUBIC assume that packet loss means congestion and halve the congestion window in response. On a wireless link, packets are frequently lost to fading, interference, or bit errors rather than queue overflow, so the algorithm cuts its rate even though the path is not actually congested, wasting capacity. A link with 1 percent random packet loss can cap a single Reno flow well below the available bandwidth. Mitigations include link-layer ARQ and hybrid ARQ to hide radio losses from TCP, explicit congestion notification so routers signal congestion without dropping, and delay-based or model-based algorithms such as BBR that estimate the bottleneck bandwidth and round-trip time directly instead of treating every loss as congestion.

Q: What is bufferbloat and how does active queue management fix it?

Bufferbloat is the high and variable latency that results when oversized router or modem buffers fill with packets during congestion. A loss-based TCP sender will keep increasing its window until a packet is dropped, so a buffer that holds several seconds of traffic will sit nearly full, adding hundreds of milliseconds to one-way delay even though throughput looks fine. Active queue management (AQM) schemes such as CoDel and PIE break this by signaling congestion early, dropping or ECN-marking packets once the standing queue delay exceeds a target (CoDel uses a 5 ms target with a 100 ms interval) rather than waiting for the buffer to fill. This keeps queues short and latency low while still allowing bursts, which is critical for real-time RF backhaul carrying voice and video.

/kuhn-JESS-chuhn kuhn-TROHL/

A family of closed-loop algorithms that regulate how fast a sender injects packets into a network so that aggregate load stays near the bottleneck capacity. By reacting to feedback signals such as packet loss, rising round-trip delay, or explicit congestion notification (ECN) marks, the sender grows and shrinks its transmission window to keep router buffers from overflowing. Effective congestion control bounds queueing delay and minimizes packet loss while still driving link utilization toward 90 to 100% of capacity, trading off raw throughput against latency. Over wireless and microwave backhaul, where radio losses are common and bandwidth varies with the channel, the choice of algorithm (Reno, CUBIC, BBR) and the use of active queue management directly govern delivered performance.

Category: Network & Telecom

Control loop: Window-based (AIMD)

Target utilization: 90 to 100% of link

How Senders Sense and Respond to a Saturated Link

Congestion control is fundamentally a distributed control problem: thousands of independent senders share a network with no central coordinator, yet they must collectively keep the offered load near each bottleneck's capacity. The classic solution, embodied in TCP, is the additive-increase/multiplicative-decrease (AIMD) loop. Each acknowledged round trip the sender raises its congestion window (cwnd) by roughly one segment; on detecting a loss it halves the window. This sawtooth probes for spare bandwidth gently while backing off sharply when the network signals trouble, and it can be shown to converge to a fair, stable equilibrium when many flows share a link.

The trigger for that back-off matters enormously over RF links. Loss-based algorithms (Reno, NewReno, CUBIC) infer congestion from dropped packets, which works on fiber where loss almost always means a full queue. On a microwave or cellular backhaul span, fading and interference drop packets even when no router is congested, so a loss-based flow needlessly halves its rate and underutilizes the link. Delay-based and model-based schemes such as BBR sidestep this by estimating the bottleneck bandwidth and minimum round-trip time directly, pacing packets at the measured delivery rate rather than filling buffers until something is lost.

The other half of the system lives in the routers. Instead of letting a tail-drop buffer fill to the brim (the cause of bufferbloat), active queue management (AQM) algorithms like CoDel and PIE watch the standing queue delay and signal congestion early, either dropping a packet or setting an ECN bit once delay exceeds a small target. Pairing a modern endpoint algorithm with AQM at the radio aggregation point keeps queues short, which is what preserves low latency for voice, video, and control traffic riding over the same backhaul.

Governing Equations

AIMD window update (per RTT):
Increase: cwnd ← cwnd + 1 | Decrease on loss: cwnd ← β × cwnd (β ≈ 0.5 Reno, 0.7 CUBIC)

Steady-state throughput (Mathis model):
T ≈ (MSS / RTT) × (C / √p) bytes/s

Bandwidth-delay product (in-flight target):
BDP = BW × RTT_min → optimal cwnd ≈ BDP

Where MSS = maximum segment size, RTT = round-trip time, p = loss probability, C ≈ 1.22, BW = bottleneck bandwidth. Example: 100 Mbps link, RTT_min = 20 ms → BDP ≈ 250 kB; at p = 10⁻⁴, MSS = 1460 B, a single Reno flow on a 50 ms path averages near 29 Mbps.

Congestion Control Algorithm Comparison

Algorithm	Signal Used	Window Decrease	Behavior on Wireless Loss	Best Use
Reno / NewReno	Triple-dup ACK, timeout	× 0.5	Over-reacts (halves)	Legacy, low-BDP paths
CUBIC	Packet loss (cubic growth)	× 0.7	Over-reacts, but recovers fast	High-BDP fiber/LTE default
BBR	Bottleneck BW + RTT model	Rate-based, no halving	Tolerant (loss not a signal)	Lossy wireless, satellite
Vegas	RTT increase (delay)	Proportional to queue	Tolerant	Low-latency, controlled nets
DCTCP	ECN mark fraction	Proportional to marks	N/A (data center)	Data-center fabrics

Common Questions

Frequently Asked Questions

How does TCP congestion control differ from flow control?

Flow control protects the receiver via the advertised window (rwnd), preventing a fast sender from overrunning a slow application's buffer. Congestion control protects the network in between via the congestion window (cwnd), the sender's estimate of how much data the path can hold without overflowing router queues. The sender transmits min(cwnd, rwnd); cwnd is adjusted from feedback such as duplicate ACKs, timeouts, or ECN marks, while rwnd is simply reported by the receiver. On a congested wireless backhaul, cwnd is almost always the limit.

Why does TCP congestion control behave poorly over wireless links?

Loss-based algorithms like Reno and CUBIC assume loss means congestion and halve the window. On a radio link, packets are often lost to fading, interference, or bit errors rather than queue overflow, so the sender cuts its rate even though the path is not congested. A link with 1% random loss can cap a single Reno flow far below the available bandwidth. Fixes include link-layer ARQ/HARQ to hide radio losses, ECN so routers mark instead of drop, and model-based algorithms such as BBR that estimate bottleneck bandwidth and RTT directly.

What is bufferbloat and how does active queue management fix it?

Bufferbloat is the high, variable latency that appears when oversized buffers fill during congestion. A loss-based sender keeps growing its window until a drop occurs, so a buffer holding seconds of traffic sits nearly full, adding hundreds of milliseconds of delay even when throughput looks fine. AQM schemes such as CoDel and PIE break this by signaling congestion early, dropping or ECN-marking once standing queue delay exceeds a target (CoDel uses 5 ms over a 100 ms interval). Queues stay short and latency low, which is critical for real-time traffic on RF backhaul.

Related Terms

← Congestion Conical Horn →