Daily visitors do not determine server CPU or memory requirements. A cached page, dynamic checkout and large report have different work per request. This tool uses a candidate instance’s measured sustainable throughput.
- Peak requests/second = requests/day ÷ 86,400 × peak-to-average multiplier.
- Mean in-flight requests at that rate = peak requests/second × mean request duration in seconds.
- Candidate instances = ceiling of
peak requests/second ÷ (tested sustainable requests/second per instance × target capacity share). - Monthly quote total = candidate instances × quoted monthly instance price + other monthly costs.
The concurrency relationship is Little’s Law (L = λW) applied to a steady operating interval. It estimates mean in-flight requests, not simultaneous logged-in users. Both rate and duration must describe the same workload and period. Daily averages multiplied by your estimate do not reveal short bursts or tail latency.
The 70% starting capacity share is an editable planning allowance, not a universal utilization threshold. The model assumes identical instances and linear scale-out, and uses a minimum of one instance. Database bottlenecks, shared dependencies, quotas, session state, failure capacity, redundancy, deployment headroom and availability-zone loss require separate tests. No CPU/RAM tier or live provider price is inferred.
Reproducible example
864,000 requests/day gives 10 requests/second on average. A 3× peak is 30 requests/second. At mean duration 0.2 seconds, mean in-flight requests are 6.
Each candidate instance tested at 25 requests/second, used at a chosen 60% share, contributes 15 requests/second. ceil(30 / 15) = 2 instances. At an assumed €50 each plus €100 other monthly costs, total is €200/month. This is a planning scenario, not a validated production architecture.
Sources
AWS Well-Architected load-testing guidance recommends testing the complete workload under realistic conditions. MIT OpenCourseWare queueing notes explain arrival rates, service times and queueing relationships. Validate the plan with production-like load tests.
Method and factual claims reviewed 16 October 2026. Sources support the principles described, not example inputs, price quotes or outcome predictions.