VPN server capacity planning

Network ceiling

Measure real sustained ingress and egress from the VPS region. The advertised port speed is not a guarantee of international throughput at peak time.

CPU and packet rate

Encryption, userspace proxy protocols, packet size, NAT, firewall rules, and traffic accounting affect CPU differently. Watch per-core saturation, not only total CPU.

Concurrency and behavior

Streaming, downloads, video calls, background messaging, idle mobile tunnels, and speed tests create different peaks even at the same customer count.

A simple first estimate

Convert the customer model into a peak load

Step	What to estimate
Step 1	Estimate the number of customers online in the busiest hour, using your own analytics as soon as they exist
Step 2	Measure average and upper-percentile throughput per online customer for the market you serve
Step 3	Multiply concurrent customers by expected throughput, then compare both directions with measured node capacity
Step 4	Reserve headroom for bursts, updates, monitoring, routing changes, and a node failure that shifts customers elsewhere

Example: 250 concurrent customers averaging 2 Mbps require roughly 500 Mbps before burst and failover margin. This arithmetic is only a starting point. Packet rate, CPU, protocol, route quality, and the distribution of heavy users still need load testing.

Test like an operator

Measure the failure boundary before customers find it

Test the actual region

Run sustained tests to destinations representative of the customer base. A benchmark between two servers in the same data center says little about the path customers use.

Test the actual protocol

WireGuard, AmneziaWG, Hysteria2, TUIC, and VLESS-Reality do not have identical CPU, transport, or client behavior. Benchmark the configuration you will sell.

Test normal and degraded states

Observe service during a node restart, a full disk, packet loss, DNS failure, a saturated port, and migration of customers to another location.

Set an operating threshold

Add or split capacity before the theoretical maximum. A service that remains responsive during bursts is more valuable than one that advertises the largest number.