2.4Performance Evaluation and Reliability

AP requires more than a qualitative sense that "availability is high" or "response is fast"—it requires the ability to estimate these with formulas and back design decisions with evidence. Performance estimation via queueing theory, and availability calculations for configurations that mix series and parallel components, are unavoidable both on the exam and in practice.

2.4.1Performance estimation via queueing theory

• Queueing theory estimates wait time and queue length under the assumption of a single server (window), random arrivals, and random service times (the M/M/1 model). Utilization (rho, ρ) = arrival rate (lambda, λ) / service rate (mu, μ) measures how much of the time the server is busy. As utilization approaches 1 (100%), wait time grows sharply and non-linearly.
• Under the M/M/1 model, the average number of customers in the system (queued plus in service) is L = ρ / (1 - ρ), and using Little's law (L = λ x W), the average time spent in the system can be derived as W = L / λ. The key property this formula reveals is that as utilization approaches 1, the denominator (1 - ρ) approaches 0, and wait time deteriorates sharply—not linearly. Pushing a server's utilization from 70% up to 90% risks a sharp spike in wait time, and needs to be approached with caution.

2.4.2Series, parallel, and mixed availability calculations

Let us trace a queueing-theory calculation with concrete numbers. For a request-processing system with an arrival rate λ = 6 requests per minute and a service rate μ = 8 requests per minute (a capacity of processing one request in an average of 7.5 seconds), utilization is ρ = λ / μ = 6 / 8 = 0.75 (75%). The average number of customers in the system is then L = ρ / (1 - ρ) = 0.75 / (1 - 0.75) = 0.75 / 0.25 = 3 (queued plus in service combined), and by Little's law the average time spent in the system is W = L / λ = 3 / 6 = 0.5 minutes (30 seconds). If the arrival rate rises to 7 per minute, ρ = 7 / 8 = 0.875, L jumps sharply to 0.875 / 0.125 = 7, and W deteriorates substantially as well. The fact that a mere ~17% increase in utilization (from 0.75 to 0.875) more than doubles the average number in the system, from 3 to 7, quantitatively demonstrates just how risky it is to operate with little headroom at high utilization. Now consider a mixed availability calculation. Suppose a system has two devices with 0.9 availability each connected in parallel (the whole works if at least one is up), and that parallel block is then connected in series with one more device with 0.95 availability. First, the parallel block's availability is 1 - (1-0.9) x (1-0.9) = 1 - 0.1 x 0.1 = 1 - 0.01 = 0.99. Treating this parallel block as a single device (availability 0.99) and multiplying it with the series device (availability 0.95) gives overall system availability = 0.99 x 0.95 = 0.9405 (about 94.05%). The key procedural point for a mixed configuration is to first combine the parallel portion into a single availability figure, then multiply it against the series portion.

2.4.3RASIS and capacity management

• RASIS (Reliability, Availability, Serviceability, Integrity, Security) should not be measured in isolation—you need to understand that its properties exist in a tradeoff relationship. For example, redundancy (a parallel configuration) added to raise availability increases cost and configuration complexity, and can lower serviceability (ease of modification). The essence of design is not maximizing each RASIS property independently, but striking a balance suited to the system's requirements.
• Capacity management is the activity of continuously monitoring and planning so that a system's processing capacity (CPU, memory, storage, bandwidth) keeps meeting requirements as future load grows. Forecasting future demand from past utilization trends, and factoring in the queueing-theory insight that wait time deteriorates sharply as utilization rises, the practical point is to judge the timing of capacity increases before utilization reaches a dangerous zone.

2.4.4Section summary

• Queueing theory (M/M/1): utilization ρ = λ / μ; average number in system L = ρ / (1 - ρ). Wait time deteriorates sharply as utilization approaches 1
• Series availability = the product; parallel availability = 1 - (product of unavailabilities). For mixed configurations, combine the parallel portion first, then compute the series step
• RASIS properties exist in a tradeoff relationship. Capacity management hinges on judging capacity increases before utilization reaches a dangerous zone