In 1-persistent CSMA, stations that find the channel busy persistently monitor it, waiting to pounce the instant it becomes idle. This aggressive behavior causes a specific problem: when multiple stations are waiting, they all transmit simultaneously, guaranteeing a collision.

Non-persistent CSMA takes the opposite approach. When a station senses a busy channel, it backs off immediately, waits for a random time period, and then senses again. This seemingly simple change has profound implications for network behavior—it breaks the synchronization that causes waiting-station collisions, at the cost of sometimes leaving the channel idle when stations have data to send.

Recall the fundamental problem with 1-persistent CSMA: if N stations are waiting for a busy channel, all N transmit at the exact same moment when the channel clears. This synchronization is the root cause of waiting-station collisions.

The solution: Don't let stations wait at the channel. Instead, when a station finds the channel busy, make it:

1. Leave — Stop monitoring the channel
2. Wait — Pause for a random duration
3. Return — Sense the channel again

By sending waiting stations away with random delays, they return at different times. Even if the channel clears while they're away, only the first to return will find it idle.

Let's formalize the non-persistent CSMA algorithm precisely.

Critical difference from 1-persistent:

Notice step 4: when the channel is busy, the station does not continue sensing. It backs off completely, waits blindly for the random duration, and only then checks the channel again. During the backoff period, the channel might become idle, but the station won't know—it's not listening.

The backoff distribution:

The random backoff delay is typically chosen from a uniform or exponential distribution:

• Uniform distribution: Backoff = Random(0, W) where W is the backoff window
• Exponential distribution: P(backoff > t) = e^(-λt)

The choice of distribution and parameters affects performance, but the key property is randomness—ensuring different stations back off for different durations.

Let's trace through several scenarios to understand non-persistent behavior in practice.

Scenario 1: Single station, channel idle

Identical to 1-persistent: when the channel is idle, transmit immediately.

Scenario 2: Three stations, channel busy, different backoffs

Stations A, B, C all have frames. Station X is currently transmitting (finishes at t=1000μs).

Collision still occurs! Both A and B happened to check at the same moment. However, notice:

• Only 2 stations collided (not all 3)
• The probability of collision is reduced compared to 1-persistent

Scenario 3: Three stations, staggered returns (no collision)

Same setup, but random backoffs spread stations out:

All three transmit successfully! Random backoffs spread stations in time, preventing collision.

Let's analyze the throughput of non-persistent CSMA mathematically, using the same model as before:

• G = Offered load
• S = Throughput
• a = τ/T = Propagation delay / Transmission time

Throughput formula for Non-Persistent CSMA:

$$S = \frac{G \cdot e^{-aG}}{G(1 + 2a) + e^{-aG}}$$

This formula is simpler than the 1-persistent case because non-persistent introduces independence between transmission attempts.

When a → 0 (negligible propagation delay):

$$S = \frac{G}{1 + G}$$

This gives maximum throughput as G → ∞: $$S_{max} = 1 \text{ (100%)}$$

Theoretically, non-persistent CSMA can approach 100% utilization! In practice, a > 0, so the actual maximum is lower.

When a is small (typical LAN):

The maximum throughput is approximately: $$S_{max} \approx \frac{1}{1 + 2a}$$

For a = 0.01: S_max ≈ 98%

Intuition for the crossover:

At low load (G << 1):

• Channel is usually idle when frames arrive
• 1-persistent: Transmit immediately → Great!
• Non-persistent: Transmit immediately → Same!
• But non-persistent sometimes backs off when it didn't need to, wasting time
• 1-persistent wins

At high load (G >> 1):

• Many stations waiting; channel rarely idle for long
• 1-persistent: All waiting stations collide immediately → Waste!
• Non-persistent: Stations spread out, fewer collisions → Better!
• Non-persistent wins

Non-persistent CSMA pays a price for its collision reduction: unnecessary channel idle time. Let's understand this tradeoff in detail.

The idle-time problem:

Consider a station A that senses the channel busy at time t=0 and backs off for 500μs. The busy transmission ends at t=100μs. The channel is now idle, but A is still in its backoff period—not sensing!

For 400μs, the channel was idle but no station was using it—even though A had data ready to send!

Quantifying the waste:

In non-persistent CSMA, the average idle time after a transmission depends on:

1. Number of backed-off stations: More stations → shorter average until one checks
2. Backoff distribution: Longer backoffs → more idle time
3. Propagation delay: Adds minimum idle time for sensing

Expected idle period ≈ (Average backoff) / (Number of waiting stations)

With few stations and long backoffs, the channel can sit idle for significant periods despite pending traffic.

The fundamental tradeoff matrix:

Let's systematically compare 1-persistent and non-persistent CSMA across multiple dimensions.

Use case guidance:

Choose 1-Persistent when:

• Network load is expected to be light (< 30% utilization)
• Latency is critical and traffic is bursty
• Implementation simplicity is valued
• Combined with collision detection (CSMA/CD) for recovery

Choose Non-Persistent when:

• Network load is expected to be high or unpredictable
• Stability under load is more important than peak performance
• Collision detection is not available (e.g., wireless)
• Traffic patterns include many simultaneous transmitters

Let's understand the intuition behind the non-persistent throughput formula. While a rigorous derivation requires queueing theory, we can build intuition.

Model assumptions:

1. Stations generate frames according to a Poisson process with rate G
2. When sensing busy, stations reschedule after random delay
3. All stations can sense all other stations (no hidden terminals)
4. Frame transmission time = 1 unit (normalized)
5. Propagation delay = a units

Time decomposition:

We can decompose time into three types of periods:

1. Successful transmission period: Duration = 1 + a (transmission + propagation)
2. Collision period: Duration ≈ 1 + 2a (overlapping transmissions + propagation)
3. Idle period: Variable duration based on backoffs

For non-persistent CSMA:

The key insight is that reschedules after sensing busy effectively thin the Poisson arrival process. If the original rate is G, the effective rate of transmission attempts becomes approximately G/(1+G) under high load.

Simplified derivation (a → 0):

When propagation delay is negligible:

• Probability of successful transmission at time t given channel sensed idle: P(success) = e^(-2Gτ) ≈ e^0 = 1 when a → 0

• But stations back off when busy, so effective arrival rate is reduced

• Throughput S = (Rate of transmissions) × P(success) × (1 transmission time)

This leads to: S = G / (1 + G)

As G → ∞, S → 1 (100% utilization in theory)

Non-persistent CSMA takes a patient approach to channel access: when the channel is busy, back off for a random duration rather than waiting continuously. This breaks the synchronization that causes waiting-station collisions in 1-persistent CSMA.

What's next:

We've seen two extremes:

• 1-persistent: Always transmit immediately when idle (probability 1)
• Non-persistent: Back off when busy, without persistent monitoring

Is there a middle ground? p-persistent CSMA offers exactly that: when sensing idle, transmit with probability p (not 1). This adds randomization even for idle-channel transmissions, further reducing collisions while avoiding excessive backoffs.

The next page explores p-persistent CSMA: its algorithm, optimal choice of p, and the elegant mathematical framework it provides.