1 Concept and purpose

Congestion control refers to the methods used in communication networks to prevent traffic from exceeding the network’s carrying capacity. When too much data is injected into shared links, routers, switches, and buffers can become overloaded, leading to delays and packet loss. Congestion control adjusts sending behavior or resource allocation so that data transfer remains efficient and stable under changing conditions.

1.1 Definition

In networking, congestion control is the coordinated regulation of packet transmission in response to signs that the network is becoming crowded. It differs from simple error handling: the aim is not only to recover from lost packets, but also to avoid creating excessive load in the first place. The term is used for both end-host protocols, such as transport-layer algorithms, and network-side mechanisms, such as queue management and traffic shaping.

1.2 Network congestion

Network congestion occurs when demand for transmission exceeds available bandwidth, buffer space, or processing capacity. It is most common in packet-switched systems, where many users share the same physical infrastructure. Congestion may be transient, lasting only briefly during bursts of traffic, or persistent, when offered load remains above capacity for longer periods.

1.2.1 Causes of congestion

Congestion can arise from several conditions. A sudden increase in traffic, such as many flows starting at once, can fill queues rapidly. Limited link capacity may create bottlenecks where many streams converge. Small buffers, slow processing at routers, and imbalanced traffic patterns can also contribute. In some cases, control algorithms themselves can interact poorly, causing oscillations or repeated bursts.

1.2.2 Effects on performance

When a network is congested, packets may be delayed in queues, dropped when buffers overflow, or delivered out of order. These effects reduce throughput and increase latency. Interactive applications can suffer from jitter, while bulk transfers may slow dramatically. Severe congestion can also trigger retransmissions, which add more load and can further worsen the condition.

1.3 Goals of congestion control

The main goals are to maintain high useful throughput, limit delay, avoid unnecessary packet loss, and distribute network capacity fairly among competing flows. A well-designed scheme should respond quickly to overload, remain stable under load changes, and make efficient use of available resources without driving the system into collapse.

2 Historical development

Congestion control emerged as a practical necessity as computer networks grew from small experimental systems into large shared infrastructures. Early experience showed that simply increasing transmission rates could reduce overall performance when buffers and links were overloaded. This led to protocols and operating principles that treated congestion as a dynamic control problem.

2.1 Early network congestion issues

In early data networks, traffic surges often caused queue buildup and heavy packet loss. Designers observed that retransmissions after loss could worsen congestion, creating feedback loops in which more recovery traffic was injected into already saturated paths. These observations helped establish congestion as a distinct networking issue, separate from link errors or routing failures.

2.2 Evolution in packet-switched networks

Packet-switched networks made congestion control especially important because multiple flows shared limited resources. As internetworks expanded, transport protocols began to include adaptive sending behavior based on inferred network conditions. Over time, congestion control moved from simple rate restraint to more sophisticated algorithms that reacted to loss, delay, and explicit feedback.

2.3 Standardization and protocol design

Standardization played a major role in spreading common congestion-control practices. Protocol specifications defined how senders should probe for available capacity, reduce load after congestion, and recover transmission efficiency. This helped create interoperability across heterogeneous devices and networks, while also giving implementers a framework for tuning performance.

3 Types of congestion control

Congestion control methods are often grouped into open-loop and closed-loop approaches. Open-loop methods try to prevent overload before it occurs by regulating traffic entry or shaping sending patterns. Closed-loop methods observe network behavior and adjust in response to feedback from loss, delay, or explicit signaling.

3.1 Open-loop congestion control

Open-loop control relies on predefined rules rather than real-time feedback from the network. Its purpose is to keep traffic within limits from the outset. This approach is often used at network boundaries or in administrative policy enforcement.

3.1.1 Admission control

Admission control decides whether new flows or sessions may enter the network. If existing load is already high, additional traffic may be delayed or rejected to preserve service quality for current users. This technique is common in managed networks and systems that need predictable performance.

3.1.2 Traffic shaping

Traffic shaping smooths bursts by controlling when packets are released into the network. By spacing transmissions more evenly, it can reduce the chance that a short spike will overflow buffers downstream. Common shaping techniques include pacing and rate limiting.

3.2 Closed-loop congestion control

Closed-loop control depends on observations from the network itself. The sender changes behavior after detecting signs of overload, such as packet loss, increasing delay, or explicit congestion messages. This makes the system adaptive, though it can also introduce measurement lag.

3.2.1 Feedback-based control

Feedback-based methods use signals that indicate current network conditions. A receiver, router, or transport endpoint may report congestion indirectly through acknowledgments, delay measurements, or control messages. The sender then lowers its sending rate or window size and later increases it cautiously.

3.2.2 End-to-end control

End-to-end control places most responsibility on the communicating hosts rather than the network core. The sender infers congestion from the behavior of packets in transit and adjusts accordingly. This model scales well because it avoids requiring every router to maintain detailed per-flow state.

3.2.3 Network-assisted control

In network-assisted schemes, routers or switches provide explicit information about congestion. This may take the form of marked packets, control bits, or rate suggestions. Such mechanisms can improve accuracy because the network can report overload before significant loss occurs.

4 Congestion control in transport protocols

Transport protocols are central to congestion control because they govern how application data is injected into the network. Their algorithms determine how rapidly senders expand traffic, how they react to loss or delay, and how they recover after reducing load.

4.1 TCP congestion control

Transmission Control Protocol implementations are widely associated with classic congestion-control behavior. TCP dynamically changes its sending window in response to acknowledgments and signals of congestion. Its design balances caution with gradual probing of available capacity.

4.1.1 Slow start

Slow start allows a sender to begin conservatively and then increase transmission rapidly as acknowledgments arrive. Despite the name, it can grow quickly during the early phase of a connection. The purpose is to discover available bandwidth without overwhelming the path immediately.

4.1.2 Congestion avoidance

Congestion avoidance takes over after initial probing. The sender increases its rate more carefully, usually in smaller steps, to avoid triggering overload. This phase aims to keep throughput high while preventing unstable growth.

4.1.3 Fast retransmit

Fast retransmit reduces waiting time when loss appears likely. If multiple duplicate acknowledgments indicate that a packet was probably lost, the sender retransmits it without waiting for a timeout. This helps restore data flow more quickly.

4.1.4 Fast recovery

Fast recovery is used after certain loss events to resume transmission without returning completely to the earliest phase of growth. It moderates the sender’s response so the connection can continue while still acknowledging that congestion has occurred.

4.2 QUIC congestion control

QUIC includes congestion-control behavior in user-space transport over UDP. Its algorithms are conceptually similar to those used in other modern transport protocols, but they are integrated with QUIC’s connection management and encryption design. This gives implementers flexibility in deploying updated control logic.

4.3 Comparison of transport-layer approaches

Transport-layer approaches differ in how they detect congestion, how aggressively they probe for unused capacity, and how they recover from overload. Window-based schemes track how much data may be in flight, while rate-based schemes focus on bytes sent per unit time. Some methods react mainly to loss, whereas others use delay or explicit signals. The choice affects responsiveness, fairness, and stability.

5 Algorithms and mechanisms

Congestion-control algorithms translate abstract goals into operational rules. They determine how much data a sender may transmit, how quickly it may increase load, and what measurements should trigger backoff or recovery.

5.1 Window-based control

Window-based control limits the amount of unacknowledged data that can be outstanding. This is a common strategy in transport protocols because it aligns sending behavior with feedback from the receiver and the network.

5.1.1 Congestion window

The congestion window is a sender-managed limit on how much data may be in flight before acknowledgment. It expands when the network appears uncongested and contracts when congestion is detected. This variable is central to many transport algorithms.

5.1.2 Receive window interaction

The receive window represents the receiver’s available buffer space. Effective transmission is limited by both the congestion window and the receive window, since data cannot be sent faster than either the network or the receiver can handle. Their interaction helps prevent receiver overload as well as network overload.

5.2 Rate-based control

Rate-based control regulates the amount of traffic sent per second rather than the number of outstanding packets. This can be useful where smooth pacing is preferred or where applications need a more direct speed limit.

5.2.1 Sending rate adjustment

Sending rate adjustment changes transmission speed according to observed conditions. If congestion increases, the rate is reduced; if capacity appears available, it may be raised cautiously. This approach is often paired with pacing to avoid short bursts.

5.2.2 Token and leaky bucket models

Token bucket and leaky bucket models are classic traffic-regulation tools. A token bucket allows bursts up to a stored credit limit while enforcing an average rate over time. A leaky bucket releases traffic at a steadier pace, helping to smooth irregular input.

5.3 Delay-based control

Delay-based methods monitor increasing round-trip time or queueing delay as an early sign of congestion. Because they can react before loss occurs, they may reduce packet drops and improve latency. Their effectiveness depends on accurate delay measurement and on distinguishing congestion from other sources of delay variation.

5.4 Loss-based control

Loss-based control treats packet loss as a signal that the network has become overloaded. This style is widely used because loss is easy to detect and often correlates with buffer overflow. However, it may respond only after queues have already filled, which can increase latency.

5.5 Explicit signaling

Explicit signaling uses direct messages or packet markings to indicate congestion. This can give senders clearer and earlier feedback than inference from loss alone. Such mechanisms are especially useful in managed environments where network devices can cooperate with endpoints.

6 Active queue management and buffering

Queue management determines how packets are stored, delayed, or dropped when links become busy. Buffering can absorb short traffic bursts, but excessive buffering may increase delay and make congestion less visible to endpoints.

6.1 Queue management

Queue management policies decide when packets are admitted to a buffer and when they are discarded or marked. Good queue management aims to prevent persistent queue buildup while keeping links efficiently utilized.

6.1.1 Tail drop

Tail drop discards incoming packets only when a queue is full. It is simple and widely implemented, but it can create synchronized losses across multiple flows, which may lead to abrupt rate reductions and inefficient behavior.

6.1.2 Random early detection

Random early detection introduces packet drops or marks before a queue reaches capacity. By providing earlier warning, it can encourage senders to slow down gradually rather than all at once. Its effectiveness depends on parameter settings and traffic patterns.

6.2 Bufferbloat

Bufferbloat is the condition in which oversized buffers cause excessive queueing delay. While large buffers can reduce packet loss, they may also keep packets waiting far too long, harming interactive traffic and masking congestion signals. Managing buffer size and queue discipline is therefore an important design concern.

6.3 Scheduling and queue disciplines

Scheduling determines the order in which packets leave a queue. Different disciplines can prioritize latency-sensitive traffic, separate flows more fairly, or shape service across classes. Scheduling interacts closely with congestion control because it influences who experiences delay and loss first.

7 Internet and network-layer approaches

Although congestion control is often associated with transport protocols, network-layer mechanisms also play a major role. Routers and switches can detect overload, mark packets, and cooperate with end hosts to improve overall behavior.

7.1 Router-based congestion signals

Routers can signal congestion through packet drops, explicit markings, or queue-related feedback. These signals inform endpoints that a path is becoming crowded. Because routers observe traffic directly, they can sometimes detect problems earlier than endpoints can infer them.

Explicit Congestion Notification allows packets to be marked instead of dropped when congestion is detected. The receiver then informs the sender, which can reduce its sending rate. This approach can preserve data while still communicating overload, provided both endpoints and network devices support the mechanism.

7.3 Quality of service interactions

Quality of service mechanisms classify traffic and assign different treatment to different packet groups. Congestion control interacts with these policies by determining how flows respond within each class. Managed priority can improve important services, but it must be configured carefully to avoid starving other traffic.

8 Performance analysis

Evaluating congestion control requires multiple metrics, since no single measure captures all aspects of network behavior. A scheme that improves throughput may raise delay, while one that reduces loss may lower utilization. Analysis therefore examines trade-offs among several performance dimensions.

8.1 Throughput

Throughput measures how much useful data is delivered over time. Effective congestion control should sustain high throughput without causing instability or excessive retransmission. For long-lived transfers, throughput is often a primary indicator of success.

8.2 Latency and jitter

Latency is the time required for a packet to travel through the network, and jitter is the variation in that delay. Congestion usually increases both, especially when queues grow. Low-latency services depend on control methods that keep buffering under control.

8.3 Fairness

Fairness concerns how bandwidth is shared among competing flows. A congestion-control system is often expected to prevent one connection from dominating a bottleneck. Fairness is not always equal division; in some settings it means allocating resources in proportion to policy or flow type.

8.4 Stability and responsiveness

Stability refers to the ability of a control system to avoid oscillation, collapse, or repeated overshoot. Responsiveness is the speed with which it adapts to changing conditions. Well-designed congestion control balances the two, since overly cautious algorithms may underuse capacity, while overly aggressive ones may create repeated overload.

9 Implementation and tuning

Practical congestion control depends on parameter choices, observation tools, and deployment details. Theoretical algorithms must be adapted to real devices, operating systems, and traffic mixes.

9.1 Parameter selection

Parameters such as window growth rate, queue thresholds, pacing intervals, and backoff factors influence behavior significantly. If settings are too aggressive, congestion may recur; if too conservative, capacity may remain unused. Proper tuning often depends on link speed, buffer size, and expected traffic patterns.

9.2 Monitoring and diagnostics

Operators use monitoring to detect congestion through metrics such as loss rate, delay, queue occupancy, and retransmission frequency. Diagnostic tools help identify bottlenecks and reveal whether a problem is caused by overload, misconfiguration, or physical link limitations. Continuous observation is especially useful in large or dynamic networks.

9.3 Deployment in operating systems and devices

Congestion-control behavior is implemented in operating-system network stacks, routers, switches, and embedded devices. Different environments may use different defaults or support different algorithms. Implementation details such as timer accuracy, pacing support, and buffer handling can materially affect performance.

10 Applications and use cases

Congestion control is relevant across nearly all modern network types, but its priorities vary by environment. Some systems emphasize low latency, while others focus on bulk throughput or predictable service.

10.1 Wired networks

In wired networks, congestion control helps manage shared links and prevent overloading access points, routers, and backbone segments. High-capacity links can still experience congestion when many users or flows converge on the same path. Control mechanisms here often focus on efficient sharing and stable queue behavior.

10.2 Wireless networks

Wireless links add variability because capacity can change with signal quality, interference, and mobility. Congestion-control algorithms in these environments must distinguish between overload and radio-related losses. This makes adaptation more complex than in many wired settings.

10.3 Data centers

Data center networks often carry large volumes of short, bursty traffic between servers. Congestion control in this setting aims to reduce queue buildup, preserve low latency, and avoid bottlenecks at top-of-rack or aggregation switches. Fine-grained pacing and explicit feedback are particularly valuable.

10.4 Real-time multimedia communication

Voice, video, and interactive conferencing require low delay and low jitter more than maximal throughput. Congestion control in these applications often tries to maintain smooth media delivery by adjusting bitrate, resolution, or packet pacing. Fast reaction to changing conditions is essential for preserving call quality.

11 Challenges and limitations

Despite many advances, congestion control remains difficult to perfect because network conditions are heterogeneous and constantly changing. Algorithms must operate across diverse links, devices, and traffic patterns.

11.1 High-speed networks

At very high speeds, packets arrive so quickly that traditional feedback loops may be too slow to react. Even small delays in measurement or signaling can produce large queues. This places pressure on algorithms to be both precise and lightweight.

11.2 Wireless loss ambiguity

In wireless systems, packet loss may result from channel errors rather than congestion. This ambiguity makes it harder for senders to interpret losses correctly. A control method that assumes every loss is congestion may reduce rate unnecessarily.

11.3 Fairness among flows

Different flows may use different algorithms, path lengths, or packet sizes, which can lead to uneven sharing. Some traffic types also have higher priority or more aggressive control behavior. Achieving fair coexistence across mixed conditions remains an ongoing concern.

11.4 Balancing utilization and delay

A network can often be driven near full utilization only by allowing some queueing, but excessive queueing increases delay. The central challenge is to keep links busy without letting buffers grow too deep. This trade-off is especially visible in latency-sensitive services.

12 Research directions

Research on congestion control continues to explore faster adaptation, better prediction, and more flexible coordination across network paths. New transport designs and deployment settings have broadened the range of possible approaches.

12.1 Machine-learning-assisted control

Machine-learning-assisted methods attempt to infer congestion patterns from network measurements and choose control actions adaptively. These approaches may identify complex relationships that traditional rules miss. Their success depends on data quality, training, and safe behavior under unfamiliar conditions.

12.2 Congestion control for multipath networking

Multipath networking allows a connection to use more than one route at the same time. Congestion control in this context must decide how to distribute traffic across paths with different capacities and delays. It also needs to avoid shifting too much traffic onto a path that is already crowded.

12.3 Congestion control in emerging transport protocols

New transport protocols continue to refine how congestion is detected and managed. Some emphasize encryption, user-space implementation, or faster deployment of updated algorithms. As network environments evolve, congestion control remains a core design element for maintaining dependable communication.