A VPN server can have a 10Gbps port and still deliver much less than 10Gbps of useful traffic. Encryption, packet processing, CPU allocation, and the network path all affect the result. So does the way people use the service.
The main 10Gbps VPN server requirements are a network path with enough sustained capacity, a CPU that can process the chosen VPN workload, enough memory for connection state, and spare capacity for traffic peaks and failures. There is no fixed core count or users-per-server number that applies to every deployment.
We size a VPN gateway from its expected workload, then test that workload on the proposed hardware. This guide explains how to do that for WireGuard and OpenVPN, including OpenVPN Data Channel Offload, or DCO.
The numerical scenarios below are planning examples. They are not Atal Networks benchmark results.
What Does “10Gbps VPN Performance” Mean?
Before comparing servers, define which rate you need to reach.
| Measurement | What it tells us | What it does not prove |
|---|---|---|
| Port speed | The nominal link capacity | Available Internet bandwidth or VPN speed |
| Plain network throughput | Performance without VPN encryption | Encrypted gateway capacity |
| Encrypted tunnel traffic | Traffic carried by the outer VPN connection | Useful application data delivered |
| Application goodput | Useful data received by the application | Performance for every packet size and user pattern |
| Sustained service capacity | Traffic supported while meeting latency and loss targets | Capacity after a node fails unless tested separately |
Goodput is the most useful starting point for user capacity. It excludes bytes that carry the data rather than form part of the application’s useful payload.
A link also carries headers and VPN overhead. As a result, a nominal 10Gbps Ethernet link cannot deliver a full 10Gbps of application payload through a VPN.
Decide whether your target means one fast tunnel, the total across many tunnels, or a busy Internet exit gateway. Those are different tests.
Test the Whole Gateway Path
An iperf3 test that ends on the VPN server measures an endpoint workload. It does not test the full forwarding path of an Internet exit service.
For a gateway test, traffic should travel from the client, through the encrypted tunnel, through the VPN server’s routing and firewall rules, and onward to a separate destination. Return traffic must follow the intended path too.
Include network address translation, or NAT, if production traffic uses it. Include the firewall rules and monitoring you plan to keep enabled.
Calculate Demand From Active Users
An account is not a traffic measurement. Neither is a configured WireGuard peer or an idle OpenVPN session.
We need to distinguish four quantities:
- Accounts: People or organizations allowed to use the service.
- Connected devices: Devices with a session or recent tunnel activity.
- Active devices: Devices moving traffic during the measurement period.
- Flows: Individual connections carried by those devices.
One active device can open hundreds of flows. A connected device can also remain almost idle.
Build a Busy-Hour Traffic Model
Start with this estimate for each user group:
Expected demand = connected devices × active share × average rate per active device
Calculate download and upload demand separately. Use rates measured during the same busy period, where possible.
| Example group | Connected devices | Active share | Download rate per active device | Estimated download demand |
|---|---|---|---|---|
| Group A | 1,000 | 25% | 12 Mbps | 3,000 Mbps |
| Group B | 200 | 50% | 20 Mbps | 2,000 Mbps |
| Total | 1,200 | Varies | Varies | 5,000 Mbps |
These rates are assumptions for the example, not standard requirements for browsing, video, or business applications.
The model is a first estimate. Compare it with total busy-hour traffic and short traffic peaks. A software release or company-wide meeting can make user activity rise at the same time.
Do not add separate groups’ percentile values and describe the result as the measured percentile for the whole service. Use an aligned traffic record to see how their peaks overlap.
Leave Spare Capacity
Suppose a candidate server delivers 8Gbps of application goodput under a representative test. That is a measured ceiling for that workload, not a recommended normal load.
If we choose a 70% operating target for this example:
8Gbps × 0.70 = 5.6Gbps operating budget
The 5Gbps example workload fits within that budget, with 0.6Gbps left before reaching the operating target. Whether that margin is enough depends on traffic variation, growth, and the recovery plan.
The 70% figure is a planning choice. It is not a universal rule.
Since the 8Gbps measurement already represents application goodput, do not subtract an estimated VPN overhead percentage from it again. Check the outer link counters separately to confirm that the physical network also has room.
Estimate Users Only After Defining the Workload
A 5.6Gbps budget could support 560 active devices averaging 10Mbps each from a bandwidth-only calculation:
5,600 Mbps ÷ 10 Mbps = 560 active devices
That does not prove the server supports 560 devices with acceptable service. Packet rate, connection state, client speed, CPU load, and traffic peaks can reduce the usable count.
It also does not tell us how many accounts we can sell. That requires measured activity patterns and a clear service policy.
WireGuard vs OpenVPN: Compare the Implementation
A protocol name alone does not describe how a server processes traffic. Record the implementation, version, operating system, and configuration used for each result.
| Test option | What to record | Main sizing question |
|---|---|---|
| Linux kernel WireGuard | Kernel version, peer count, MTU, CPU, and NIC settings | Does the intended traffic spread across available processing resources? |
| Userspace WireGuard | Exact application, version, and packet-processing settings | How does this implementation behave on the chosen host? |
| OpenVPN without DCO | Version, transport, cipher, process layout, and client count | Does a busy process or CPU core limit throughput? |
| OpenVPN with DCO | OpenVPN version, driver, cipher, supported settings, and runtime status | Is the data channel actually using offload during the test? |
WireGuard and CPU Encryption Features
WireGuard uses ChaCha20 for encryption and Poly1305 for authentication. It carries tunnel packets over UDP.
This means an AES acceleration feature is not a direct measure of WireGuard performance. CPU architecture, software implementation, packet handling, and the network path still need testing.
Do not treat all userspace WireGuard programs as equivalent. Tailscale’s published engineering work shows how changes to batching and network offloads affected its own implementation. Those results are useful technical evidence, but they are not a capacity promise for another application or server.
OpenVPN With and Without DCO
OpenVPN DCO moves data-channel processing into the kernel on supported systems. The userspace process still handles control functions.
For a fair comparison, run the intended OpenVPN deployment with DCO disabled and then with DCO active, where the configuration supports both. Keep the hardware, client workload, cipher, and network path consistent.
Installing a package with DCO support does not prove that offload is active. Check the startup logs and runtime state. Unsupported settings or a missing driver can change the result.
DCO support depends on the OpenVPN release, operating system, driver, and configuration. Consult the documentation for the release you deploy. The current project documentation lists supported authenticated-encryption ciphers and configuration limits.
Do not apply an old non-DCO benchmark to a current DCO deployment.
Match the Protocol to the Service
Throughput is one requirement. Client support, access control, authentication, network restrictions, and operational tools also affect the choice.
Test OpenVPN over TCP separately if you need that transport. Do not present a UDP result as proof of TCP performance.
For installation basics, see our VPN setup guide. Here, we focus on capacity after the required configuration is in place.
CPU, RAM, and Network Requirements
A fixed shopping list cannot prove 10Gbps VPN capacity. We use the following checks to select hardware and identify the resource that needs more capacity.
CPU: Inspect Each Core
A server can show moderate total CPU use while one core is fully occupied. That core may handle a busy process, a network queue, or another part of the packet path.
Record:
- CPU model and generation.
- Physical cores and logical threads.
- Sustained clock speed under load.
- Per-core use, including network processing.
- Thermal or power-related clock limits.
- CPU allocation and steal time on virtual machines.
Steal time is time when a virtual CPU waits while the host runs other work. A VPS with a large vCPU count can still have inconsistent processing time.
More cores help only when the workload can use them. A higher core count does not guarantee that one tunnel or one busy processing path will run faster.
Linux receive-side scaling, or RSS, can distribute incoming flows among NIC queues. Related mechanisms can move packet processing across CPUs. The distribution depends on the device, driver, and configuration.
Inspect the actual queue and CPU distribution before changing settings. Encrypted outer flows may distribute differently from the application flows inside them.
RAM: Size for State and Services
There is no fixed amount of RAM required per gigabit of VPN traffic.
Memory use depends on the operating system, peer or session state, packet buffers, firewall connection tracking, monitoring, and any other services on the gateway.
For an Internet exit server, record the number of tracked connections during the busy period. Many small application connections can create a different memory load from a few large transfers.
Watch memory pressure, swap activity, connection-tracking limits, and allocation failures. Raising a connection limit without checking memory use can move the failure elsewhere.
NIC, PCIe, and NUMA: Check the Path to the CPU
A 10Gbps label does not describe every part of the network path. Check the negotiated link speed, driver, queue count, interface errors, and PCIe connection.
On a multi-socket server, note which CPU socket connects to the NIC and where the workload runs. Memory and device access across sockets can affect performance.
Record offload settings before each benchmark. Features such as packet segmentation and receive aggregation can change CPU cost and the packet sizes visible to software. Keep these settings documented so another engineer can repeat the test.
Storage: Check Logging and Recovery Needs
Forwarding VPN traffic usually does not require writing each packet to disk. Storage can still become busy if the server records detailed flow logs or runs other services.
Test the logging level you intend to use. Measure disk latency, write rate, retention needs, and recovery procedures. A benchmark with logging disabled may not describe the production service.
Packet Size, MTU, and Traffic Direction
A server’s bandwidth result needs packet-size context. Small packets require more packet-processing work for the same payload rate.
Packet Rate Can Become the Limit
At a payload rate of 1Gbps, the approximate packet rate is:
Packets per second = payload bits per second ÷ (payload bytes per packet × 8)
| UDP payload size | Payload rate | Approximate packets per second |
|---|---|---|
| 1,200 bytes | 1 Gbps | 104,167 |
| 120 bytes | 1 Gbps | 1,041,667 |
This calculation excludes UDP, IP, VPN, Ethernet, and physical-layer overhead. It is not an Ethernet wire-rate calculation.
The smaller payload requires ten times as many packets in this example. That does not mean CPU use will rise by exactly ten times. Batching, offloads, and fixed processing costs affect the result.
Test both large transfers and small-packet traffic. A large-packet benchmark alone can hide a packet-rate limit.
MTU Must Fit the Actual Path
The maximum transmission unit, or MTU, limits packet size at a given layer. A VPN adds outer headers, so the tunnel must fit within the available path MTU.
A poorly matched MTU can cause fragmentation, loss, or connections that stall on larger transfers. IPv4 and IPv6 paths also need separate checks.
Do not copy one MTU value into every deployment. Test the intended client networks and tunnel path. For TCP, MSS adjustment may address some packet-size problems, but it does not fix all UDP traffic or every path issue.
A Full-Duplex Port Does Not Mean 20Gbps of VPN Payload
A 10Gbps full-duplex port can receive and transmit at the same time. Adding those two rates does not automatically create 20Gbps of useful VPN capacity.
Consider a single-interface Internet exit gateway:
- A user download enters from the Internet, then leaves toward the VPN client.
- A user upload enters from the VPN client, then leaves toward the Internet.
Both traffic directions consume receive and transmit capacity on that interface. Roughly, the combined download and upload payload contributes to each direction, with different overhead on the encrypted and unencrypted legs.
A single one-way transfer is not automatically limited to half the port rate either. The answer depends on which links it crosses and where traffic enters and leaves.
Measure each physical interface separately. Label forward, reverse, and bidirectional results, and state whether a reported total adds both directions.
Can a VPS Handle a 10Gbps VPN Workload?
A VPS can be suitable when its CPU and network allocation meet the workload. The product’s advertised host port speed is not enough information.
Ask the provider:
- Is 10Gbps available to this guest or only to the physical host?
- Is guest bandwidth capped, shared, or subject to a sustained-use policy?
- Are CPU resources shared or dedicated?
- Are there packet-rate limits?
- Does the guest support the required kernel, tunnel device, and DCO driver?
- What transfer allowance and overage rules apply?
- Can we run a permitted capacity test before committing?
A dedicated server gives us direct control over its physical CPU and NIC configuration. That can make troubleshooting easier. It still does not prove that the upstream network or VPN configuration will meet the target.
Nuestra 10Gbps VPS vs. dedicated server comparison explains the resource-allocation differences in more detail.
A Practical 10Gbps VPN Test Matrix
We need separate tests for network capacity, encryption, packet processing, and user behavior. One peak speed result cannot answer all four questions.
Use traffic generators and receivers with enough capacity for the target. A slow client or receiver can make a capable gateway appear slow.
| Test | Workload | What it helps identify |
|---|---|---|
| Plain network baseline | One TCP flow, then multiple flows without a VPN | Network or test-host limits |
| Routing baseline | Forward traffic through the intended firewall and NAT rules without encryption | Forwarding and policy cost |
| Single-tunnel test | One peer with one flow, then several flows | Single-flow and single-peer limits |
| Many-peer test | Independent clients or peer instances carrying traffic | Aggregate capacity across peers |
| Protocol comparison | WireGuard, OpenVPN without DCO, and confirmed OpenVPN DCO | Implementation-specific limits |
| Packet-size test | Controlled UDP payloads such as 120, 512, and 1,200 bytes | Packet-rate limits and loss |
| Direction test | Upload, download, then both directions | Directional and shared-resource limits |
| Application test | Busy-hour traffic mix with production rules and logging | Service quality under expected demand |
| Connection-change test | New sessions, rekeys, and a reconnect wave | Control-plane pressure |
| Failure test | Remove one gateway and restore service | Remaining capacity and recovery behavior |
Use larger packet sizes only after checking the path MTU. Keep client count, traffic direction, and application mix explicit in each result.
Record Enough Detail to Repeat the Test
Record hardware, operating system, kernel, VPN software, cipher, DCO state, NIC driver, MTU, and offload settings. Include the topology and the number of independent peers.
During each run, collect:
- Receiver-side goodput and offered traffic rate.
- Per-core CPU use on the gateway and test hosts.
- Interface rates, errors, and drops.
- TCP retransmissions or UDP loss and jitter.
- Latency before and during load, including p95 and p99.
- Memory use, connection state, and logging activity.
The p95 latency is the value at or below which 95% of recorded samples fall. It helps show delays that an average can hide.
Use iperf3 Without Treating Streams as Users
iperf3 supports reverse-direction tests with -R, parallel streams with -P, and bidirectional tests with –bidir. UDP tests can set an offered rate with -b and a payload size with -l. With parallel streams, the bandwidth setting applies per stream. Check the installed version’s manual before running tests.
Eight parallel streams through one VPN peer still represent one peer. They do not prove performance for eight independent users or tunnels.
Record the iperf3 version as well. Its threading behavior changed in version 3.16, so older and newer releases can place different limits on the test host.
Increase Load in Steps
Start below the expected operating rate. Raise the offered traffic in stages while watching goodput, loss, latency, and CPU load.
A workable test plan might use a warm-up period, three 60-second runs at each candidate rate, and a 30-minute sustained run near the proposed operating limit. These durations are a starting procedure, not a formal standard. Extend the runs for longer traffic cycles or intermittent problems.
Define acceptable loss, latency, and recovery time before testing. The usable capacity is the highest workload that stays within those targets with the required spare capacity. It is not necessarily the highest number seen for a few seconds.
How to Read the Results
A useful benchmark tells us what to change next.
| Observation | Possible cause | Next check |
|---|---|---|
| One CPU core stays busy while total CPU looks low | A process, queue, or packet path limits work distribution | Per-core load, process use, and queue distribution |
| Most cores stay busy, and throughput stops rising | CPU processing capacity may be exhausted | Lower offered rate and compare loss and latency |
| VPN and plain tests stop at a similar rate | Network, client, receiver, or provider limit | Test-host load and interface counters |
| Small packets fail before large packets | Packet-rate processing limit | Packets per second, drops, and network CPU work |
| Larger transfers stall while small requests work | Possible MTU or path issue | Packet sizes, path MTU, and retransmissions |
| Latency rises sharply near the speed ceiling | Queue buildup or resource saturation | Reduce load and inspect queues |
| Performance varies between VPS runs | Possible host contention or path variation | Steal time, time-of-day patterns, and plain baseline |
| Reconnects fail during a busy period | Control-plane, authentication, or dependency pressure | Session setup rate and dependency logs |
These observations are starting points, not final diagnoses. Change one factor at a time and repeat the affected test.
Size for Failures and Growth
A fleet that meets normal demand may not have enough capacity when one gateway is unavailable.
Suppose each gateway has a tested operating budget of 5.6Gbps and total demand is 8Gbps:
| Deployment | Normal combined budget | Budget after one gateway fails | Bandwidth-only result |
|---|---|---|---|
| Two gateways | 11.2Gbps | 5.6Gbps | Cannot carry the full example demand |
| Three gateways | 16.8Gbps | 11.2Gbps | Has enough remaining bandwidth in this example |
The second design works only if clients can reach the remaining gateways and the workload distributes as intended. CPU limits, packet rates, authentication services, and network paths still need testing.
A load balancer does not automatically preserve VPN tunnel state or NAT mappings. Clients may need to reconnect, and existing application connections may break. Test the actual recovery process.
Add growth based on measured busy-hour demand and the time needed to deploy more capacity. Set alerts early enough to act before service quality falls.
Choose the Server From a Written Requirement
Before ordering, prepare a short requirement that includes the protocol, active peers, expected traffic rates, packet mix, and recovery target.
| Requirement | Evidence to request or collect |
|---|---|
| CPU capacity | Workload-specific results on the proposed CPU |
| Sustained network capacity | Guest or server limit, port policy, and test results |
| Memory capacity | Peak state-table and memory use during a representative run |
| Software support | Required kernel, tunnel access, and offload support |
| Transfer costs | Included traffic, billing method, and excess-use terms |
| Service quality | Latency and loss measurements at the operating target |
| Failure capacity | Recovery test with one gateway unavailable |
For the wider purchasing checklist, read how to choose a 10Gbps dedicated server.
If you are considering our 10Gbps server options, share your VPN protocol, expected active devices, client regions, and busy-hour upload and download demand. We can use those requirements to discuss a suitable configuration and the tests needed before production use.
Preguntas frecuentes
How many CPU cores does a 10Gbps VPN server need?
There is no universal core count. The requirement depends on the CPU model, VPN implementation, packet sizes, peer count, and network configuration. Test the intended workload and inspect each core. A high total core count does not fix every single-flow or queue limit.
How many concurrent users can a 10Gbps VPN server support?
Start with active devices and their busy-hour traffic rates. Divide the tested operating budget by the expected rate per active device for a bandwidth-only estimate. Then test packet rate, connection state, latency, and reconnect behavior. Connected but idle devices are a separate part of the model.
Is WireGuard always faster than OpenVPN?
No universal speed ranking applies to every deployment. Compare the exact WireGuard implementation with the intended OpenVPN version and DCO state. Use the same hardware, network path, traffic direction, and workload. Older non-DCO results do not describe every current OpenVPN deployment.
Does AES-NI help WireGuard reach 10Gbps?
WireGuard uses ChaCha20-Poly1305 rather than AES for its data encryption. AES-NI support alone is therefore not a WireGuard sizing rule. For an OpenVPN deployment using AES, assess the CPU and software’s available acceleration through a representative test.
Does OpenVPN DCO guarantee 10Gbps?
No. DCO changes where data-channel work runs, but the CPU, NIC, clients, network path, and configuration still set limits. Confirm that DCO is active, then measure the complete gateway workload.
How much RAM does a 10Gbps VPN server need?
Port speed does not set a fixed RAM requirement. Measure peer state, firewall connection tracking, packet buffers, monitoring, and any additional services. Leave room for traffic peaks and reconnects, and check for memory pressure during sustained load.
Is a 10Gbps port the same as 10Gbps unmetered bandwidth?
No. Port speed describes link capacity. Unmetered describes a billing arrangement, subject to the plan’s terms. Confirm whether the port is shared or dedicated, what sustained traffic is allowed, and whether the provider applies other limits.
Why is one VPN client slower than the server’s total benchmark?
The client may have a slower CPU, Wi-Fi link, access connection, or network path. A single flow or tunnel may also use processing resources differently from many peers. Test individual-client performance and aggregate server capacity separately.
