Contention-Based Medium Access

Contention trades reservation overhead for collisions and variable waiting. Each station decides when to attempt transmission using local information; when attempts overlap incompatibly, affected frames fail and must be retried. ALOHA makes this trade mathematically visible. CSMA improves it by listening first, while collision detection or avoidance addresses what listening cannot know soon enough.

Pure ALOHA #

Assume every frame takes time $T$, attempted frame starts form a Poisson process, and the offered load is $G$ attempts per frame time—including retransmissions. A frame beginning at time $0$ succeeds only if no other frame begins in $(-T,+T)$: the vulnerable interval is $2T$.

For a Poisson process, the probability of zero other attempts in an interval with mean $2G$ is $e^{-2G}$. Successful throughput in frames per frame time is therefore

$$S=G e^{-2G}.$$

Differentiating,

$$\frac{dS}{dG}=e^{-2G}(1-2G)=0$$

gives $G=1/2$ and

$$S_{max}=\frac{1}{2e}\approx0.184.$$

Even at its optimum, pure ALOHA converts only about 18.4% of frame-time capacity into successful frames under these assumptions. More offered load beyond the optimum reduces useful throughput because retries create still more collisions.

Slotted ALOHA #

If all starts are restricted to slot boundaries and one frame occupies one slot, only attempts in the same slot collide. The vulnerable interval falls to $T$:

$$S=G e^{-G}, \qquad S_{max}=\frac{1}{e}\approx0.368 \quad\text{at }G=1.$$

Slot synchronization doubles the ideal maximum throughput but adds clock/distribution requirements and waiting until the next boundary.

Stability depends on retransmission control #

$G$ includes new traffic and retries. Let fresh load be $\lambda$ and retry load be $r$; then $G=\lambda+r$. If every backlogged station retries aggressively, one collision creates multiple near-simultaneous retries, which create more backlog. The system can move from a low-load stable operating point to persistent collision even while fresh traffic is below the theoretical peak.

Random backoff decorrelates retries. A useful backoff policy must answer:

• how the contention window changes after failure and success;
• whether countdown continues while the medium is busy;
• the retry limit and final failure signal;
• fairness between old backlog and new arrivals;
• how much delay tail is acceptable.

Maximum throughput alone is insufficient: a policy can achieve good aggregate throughput while starving one station or producing an operationally unacceptable access-delay tail.

Carrier sense and propagation #

CSMA listens before transmitting. If the medium is busy, a station defers; this avoids starting on top of a signal that has already arrived. It cannot prevent two distant stations from beginning within one propagation delay of each other.

Let one-way propagation be $\tau$ and frame transmission time be $T$. The ratio

$$a=\frac{\tau}{T}$$

captures how stale carrier sense can be relative to a frame. Small $a$ improves contention efficiency because stations learn of another start quickly. Higher link rate or shorter frames reduces $T$, so the same physical span becomes harder unless the collision domain is constrained.

CSMA/CD: detect and abort #

Shared half-duplex Ethernet used carrier-sense multiple access with collision detection. A transmitter monitored the medium while sending; upon collision it stopped the failed frame, signaled the event, and retried after a randomized binary exponential backoff. The minimum frame and maximum collision-domain extent ensured a transmitter was still sending when a worst-case collision propagated back.

Modern switched full-duplex Ethernet links do not contend: each direction is point-to-point, so CSMA/CD is unnecessary there. Output-queue congestion and packet loss can still occur inside a switch, but they are not link collisions.

CSMA/CA: reduce collision probability #

A Wi-Fi transmitter cannot reliably infer at its antenna that every intended receiver decoded its frame. IEEE 802.11 distributed access therefore uses collision avoidance:

1. observe the channel idle for the required interframe interval;
2. choose a random backoff counter from the current contention window;
3. decrement it only during eligible idle slots and freeze it while busy;
4. transmit when it reaches zero;
5. infer success from an acknowledgement, otherwise enlarge the contention window and retry subject to limits.

Collisions remain possible when contenders choose the same slot or are hidden from each other. RTS/CTS can reduce the cost of some hidden-terminal collisions by reserving airtime with shorter control frames, but its overhead is not automatically worthwhile for every frame.

What to measure #

• offered attempts $G$, successful throughput $S$, and fresh application load separately;
• collision or missing-ack retry counts by contention stage;
• contention-window distribution and access-delay percentiles;
• channel busy time split into successful payload, control, retransmission, and foreign energy;
• hidden-node/asymmetric-link evidence and capture effects;
• fairness per station and per unit airtime.

Random access works well when demand is sparse and failures are quickly decorrelated. As offered load rises, the mechanism must make congestion visible through backoff, delay, or admission rather than letting retries hide overload as “more traffic.”