5.2Performance & traffic design

For a network designer, it is not simply a matter of "more bandwidth is always better." You must estimate the bandwidth and device capacity that are necessary and sufficient to meet actual performance requirements (response time, voice quality, etc.), and use quantitative calculations such as queueing theory as the grounds for that estimate. This section covers what the performance metrics mean and how to translate calculation results into design decisions.

5.2.1Performance metrics (throughput, bandwidth, latency, jitter, packet loss)

• Bandwidth is the theoretical maximum data volume a link can carry (bps). Throughput is the actual effective data volume that flows, and is always less than bandwidth due to protocol overhead, congestion, or device processing capacity. Increasing bandwidth does not necessarily increase throughput proportionally—if the bottleneck lies elsewhere (e.g., device CPU), the increase has no effect.
• Latency is the time from when data is sent until it reaches the recipient. Jitter is the variation (fluctuation) in latency. For real-time communication such as voice and video, not just the absolute latency but a large jitter is perceived as audio dropouts or jerky video. Packet loss is the fraction of packets lost in transit, mainly caused by congestion or physical-layer degradation.

5.2.2Queueing theory (M/M/1)

• Queueing theory mathematically models congestion where packet arrivals and processing at a router, etc. occur randomly. The M/M/1 model—arrivals and processing times both exponentially distributed, with a single server (e.g., one router interface)—is the most basic model and underlies the performance design of network equipment.
• Utilization ρ (rho) is the fraction of processing capacity in use: ρ = λ / μ (λ = mean arrival rate, μ = mean processing capacity). Unless ρ < 1, the queue (buffer) grows without bound, overflows, and leads to packet loss.
• In the M/M/1 model, the average waiting time (time spent waiting in the queue) is Wq = ρ / (μ - λ). Increasing processing capacity μ (faster equipment, more bandwidth) or decreasing arrival rate λ (distributing or reducing traffic) is the basic way to lower ρ and Wq.

Suppose a router linking two sites sees a mean packet arrival rate of λ=800 packets/sec and a mean processing capacity of μ=1000 packets/sec. First, the utilization ρ is ρ = λ / μ = 800 / 1000 = 0.8 (80%), and since ρ < 1, the steady state is stable. Next, the average waiting time is Wq = ρ / (μ - λ) = 0.8 / (1000 - 800) = 0.8 / 200 = 0.004 sec (4 ms). Now suppose there is a plan to launch VoIP calls over this link, with a voice-quality requirement to keep the average waiting time at 2 ms (0.002 sec) or less. Solving backward for the required processing capacity μ' means solving Wq' = (λ/μ') / (μ' - λ) ≤ 0.002 using ρ' = λ / μ'. Substituting λ=800 and rearranging gives 800 / (μ'(μ' - 800)) ≤ 0.002, i.e., the quadratic inequality μ'(μ' - 800) ≥ 400000. Solving μ'^2 - 800μ' - 400000 = 0 with the quadratic formula gives μ' = (800 + √(800^2 + 4×400000)) / 2 = (800 + √2240000) / 2 ≈ (800 + 1496.7) / 2 ≈ 1148.3, showing that processing capacity must be increased from 1000 pps to roughly 1150 pps or more. This can be achieved not only by "adding more bandwidth" but by several design choices, such as upgrading to a higher-performance router or distributing routes to lower the arrival rate λ per device. Rather than simply "add more bandwidth," the practical decision is grounded in the required processing capacity derived backward from the target via queueing theory, and then choosing which means to use to achieve it.

5.2.3Traffic-volume calculation and capacity management

• Capacity management is the process of planning ahead to secure the necessary bandwidth and device performance based on current usage and projected future growth. In practice, it is common to continuously monitor peak utilization as a trend in ρ and set an operational rule to consider an upgrade once ρ exceeds a certain threshold (e.g., 0.7-0.8)—because Wq increases sharply as ρ approaches 1, waiting until after the threshold is crossed often makes remediation too late.
• Bottleneck diagnosis is the process of isolating, when performance degrades, whether the cause lies in link bandwidth, device CPU/memory, or traffic concentration from a specific application. If link utilization is low but throughput fails to grow, the device's processing capacity (μ) is likely the bottleneck, and a device upgrade or load balancing is a more effective countermeasure than increasing link speed.

5.2.4Section summary

• Queueing (M/M/1): ρ=λ/μ, Wq=ρ/(μ-λ). Solve backward from a target Wq for the required μ, then decide the means of upgrade (device replacement, distribution, etc.)
• Throughput is always less than the theoretical value of bandwidth. Isolate whether stagnation is due to the link or a bottleneck such as device CPU
• Capacity management best practice detects a ρ threshold breach in advance and upgrades proactively. Jitter is the variation in latency, not the average