The advice on this page is intended to help firewall administrators handle increased VPN volume when using pfSense® software, both in terms of throughput and number of connected users.
The advice on this page is relayed from experience and from community members. The advice on this page may not apply to all environments or use cases, and has not been definitively proven to help, but is offered in case others find it useful.
WireGuard has been removed from the base system in releases after pfSense Plus 21.02-p1 and pfSense CE 2.5.0, when it was removed from FreeBSD.
If upgrading from a version that has WireGuard active, the upgrade will abort until all WireGuard tunnels are removed. For more details, see the Release Notes
WireGuard is available as an experimental add-on package on pfSense Plus 21.05, pfSense CE 2.5.2, and later versions. The settings for the WireGuard add-on package are not compatible with the older base system configuration.
The WireGuard package is still under active development. Follow the development progress on the developer’s YouTube channel
No Artificial Limits¶
The firewall does not place any artificial limits on VPN connections. Any limitations encountered are due to settings, the hardware/environment, or the underlying technology.
IPsec is Fastest¶
IPsec is faster than other choices when configured correctly and when using hardware acceleration. If both client and server support IPsec, use IPsec. WireGuard is nearly as fast, however, and depending on the environment and availability of hardware acceleration, may be faster in some cases.
Use External Authentication¶
For user-based authentication, the most efficient method of user management for large numbers of accounts is an external authentication source, such as a RADIUS server, LDAP server, Active Directory (Via LDAP or RADIUS/NPS), etc.
If additional users are unable to connect, look in the logs on both the client and server side for specific error messages before seeking support.
Use Hardware Acceleration¶
Using a cryptographic accelerator such as a QAT, AES-NI, or SafeXcel will help greatly with throughput and crypto- related tasks.
pfSense® Plus software includes support for QAT, CESA, and SafeXcel hardware found on several Netgate Appliances. QAT is also available as an add-on card for certain models. pfSense® Plus software and pfSense CE software both include support for AES-NI.
For more information on these devices and their capabilities, including how to configure and test them, see Cryptographic Accelerator Support.
Use AES-GCM or ChaCha20-Poly1305¶
Using efficient encryption will increase security and performance. Authenticated Encryption with Additional Data (AEAD) ciphers combine encryption and authentication in one step, eliminating the need for additional hashing. AES-GCM and ChaCha20-Poly1305 are both AEAD ciphers. IPsec and OpenVPN can use AES-GCM. OpenVPN can also use ChaCha20-Poly1305, and that is also the only cipher supported by WireGuard.
Client support varies by platform.
Use Accelerated Ciphers¶
Certain hardware may accelerate ciphers so that choices are faster or more efficient. For hardware sold by Netgate, see the Netgate Appliances page for performance data and recommendations.
Disable Performance-Limiting Mitigation Settings¶
While we do not have any data on if or by how much they may impact VPNs, CPU vulnerability mitigation methods such as Kernel PTI and MDS mode can potentially degrade total performance. The potential for exploitation is minimal since arbitrary code cannot be run on the firewall except by users which already have the equivalent of administrator-level access. To ensure this risk stays low, only allow trusted administrators to access the firewall GUI and shell (SSH or console). The settings to enable/disable these features are under System > Advanced on the Miscellaneous tab.
Check Tunnel Network/Virtual Address Pool Sizes¶
Both IPsec and OpenVPN can assign addresses to clients out of a pool for remote access/mobile VPNs. The sizing of this pool limits how many clients can connect. For example, the maximum number of users in a /24 pool is 252, but other settings may reduce that value. See the sections below for more specific advice.
Use “Secure Enough” Settings¶
While we do not recommend deliberately using weak configurations, in some cases trade-offs are made for security between two secure ciphers or settings where one may offer even better security, but the lower of the two is still secure. In these cases, using the “Secure Enough” option can provide efficiency vs increased security. So long as the decision is informed, there may be some performance gained without compromising security in an unacceptable way. For example, with AES-GCM a key length of 128 bits is still considered secure. A 256 bit key is more secure, but the 256 bit key could put more of a burden on the hardware.
Consider Split Tunneling¶
Configurations which send all client data over the VPN, including Internet-bound traffic, will consume more resources than those which only send traffic for specific subnets. There are plenty of valid reasons to use either kind of configuration, however, when resources are stretched thin, easing the traffic burden on the VPN may justify switching to split tunneling rather than tunneling everything. Depending on the type of VPN and client, this may require adjustments on the server, the client, or both. See the sections below for specific recommendations.
Use Multiple Firewalls¶
In some instances, the burden may be too great for any single firewall to handle. In cases like this, multiple firewalls can be used to handle the required number of clients or throughput, at a cost of greatly increased complexity. There is no way to automatically balance between nodes in this manner, but such a configuration could be manually managed. This would also likely require the capability to have multiple external addresses on the WAN so each firewall can work in parallel, and also increases the complexity of routing on the internal side.
IPsec is well-suited to high throughput by default, especially given the advice above, but there are additional IPsec-specific tweaks which may help.
See the TNSR product page for information about using TNSR for even larger total site-to-site throughput needs.
Optimal Encryption Settings¶
Use AES-NI capable hardware.
In Phase 1 (IKE) settings, use:
AES128-GCM with 128 bit key length for the Algorithm
AES-XCBC for the hash, which in this case is effectively a Pseudo-Random Function (PRF). This will yield the highest performance in combination with AES128-GCM on hardware which can accelerate both (e.g. AES-NI)
While this may lead to higher performance when hardware is capable of accelerating this algorithm, it is less secure than other choices such as SHA256.
Using a more secure algorithm is better in most use cases. IKE negotiation does not happen frequently unless there are large quantities of tunnels or mobile clients, so it may not be critical to accelerate this function in the majority of deployments.
In Phase 2 (Child SA) settings, use:
AES128-GCM with 128 bit key length for the Algorithm
Do not select any Hash Algorithms. A hash algorithm is unnecessary for AES-GCM as it already includes authentication.
Enable Multiple Phase 1 and Phase 2 Proposals¶
Multiple Phase 1 and Phase 2 encryption proposals may be configured in the GUI. Enabling multiple combinations of settings will allow peers to choose the most optimal settings which both sides support.
Enable Asynchronous Cryptography¶
IPsec cryptography jobs can be dispatched multi-threaded to run in parallel and increase performance. However, not all platforms and configurations fully support this function. To enable this capability, check Asynchronous Cryptography under VPN > IPsec on the Advanced tab.
Be on the lookout for IPsec traffic drops/failures to pass with this setting enabled. See https://redmine.pfsense.org/issues/8964 for more information.
As mentioned above, split tunneling would only send traffic for specific subnets
across the VPN rather than sending all traffic. On IPsec, this can be done in
some cases by listing the specific networks in Phase 2 entries for the Mobile
IPsec P1 rather than
0.0.0.0/0. On the mobile clients tab, set Provide a
list of accessible networks to clients. Even with that set, certain cases such
as Windows 10 may require additional changes to direct clients to send only
specific traffic over the tunnel.
Use IPsec or WireGuard Instead¶
As mentioned previously, where possible, use IPsec or WireGuard instead. IPsec and WireGuard are much more efficiently integrated into the operating system, and are capable of much greater throughput than OpenVPN.
UDP has less overhead for tunneled data, and if a client has to retransmit, it won’t compound the problem by retransmitting both inside and outside the tunnel. Unless there are extenuating circumstances which require TCP, use UDP.
Use TLS for Authentication Only¶
OpenVPN can use TLS for both authentication and for encryption of the control channel. Performing control channel encryption adds more overhead, which can add up with many clients. If control channel encryption is not required, consider using TLS for only authentication instead. No matter which option is chosen, traffic carried by OpenVPN is encrypted.
Use a CPU with AES-NI when possible, and use AES-GCM for the Encryption Algorithm when possible. ChaCha20-Poly1305 may also give a performance boost if hardware acceleration is unavailable. Note that for AEAD ciphers such as AES-GCM and ChaCha20-Poly1305, OpenVPN ignores the setting for Auth Digest Algorithm.
For OpenVPN, AES-GCM and ChaCha20-Poly1305 can only be used in SSL/TLS mode.
Use Data Cipher Negotiation¶
Data cipher negotiation can be used to set preferences so that more efficient ciphers can be preferred by clients where possible, but others can be used when necessary. Set high-priority selections such as AES-128-GCM first, followed by others like AES-128-CBC.
As mentioned in the general section above, split tunneling only sends traffic for specific subnets across the VPN rather than sending all traffic. With OpenVPN, this can be done by Unchecking the Redirect IPv4/IPv6 Gateway option(s) and configuring IPv4/IPv6 Local Network(s) entries instead. Clients may still override this behavior remotely, however, so check the client configurations as well.
The firewall does not impose any connection limits by default, but an administrator may have chosen to configure a limit on the number of connections via the Concurrent Connections setting on servers. Ensure this is either unset or set high enough to accommodate the required number of users.
Though using compression is tempting to squeeze extra throughput out of slower links, it is both inefficient and insecure. Most data sent across VPNs in modern environments is already encrypted or otherwise uncompressible, which wastes CPU when attempting to compress. Additionally, vulnerabilities such as VORACLE can allow attackers to glean information about encrypted data when it has been compressed. Disabling encryption will mitigate that attack and also reduce CPU overhead. On the server, set Compression to Disable Compression.
Normally, if an OpenVPN client connects using the same username or certificate CN, the older connection is broken in favor of the new connection. This is more secure, but does not allow any given user to connect multiple times. Circumstances may necessitate supporting this, and in some environments it’s not possible to give every device a unique username and/or certificate. Check Duplicate Connection in the OpenVPN server settings to allow multiple connections from the same user.
OpenVPN defaults to subnet topology which uses addresses more efficiently, but
if the VPN was configured initially on older versions, or if an older guide was
followed, it may still be using net30 topology. Using a common example tunnel
10.0.8.0/24, with subnet topology, the VPN can have a maximum
of 252 users but with net30, it can only have 63. This is because in net30
mode, each user receives a /30 subnet which utilizes four IP addresses for each
user. In subnet mode, the server uses a single address and the client uses a
single address, which is much more efficient.
Use UDP Fast I/O¶
This option is experimental but for those who have used it, it can result in much higher throughput. Not all platforms support it, however.
Increase Send/Receive Buffer¶
The default buffer size is safe, but not optimal. Increasing the buffer size to 512KiB on both sides can result in greater throughput. Results will vary by platform, internet link speed, and other factors. May require experimenting with multiple values to find the most efficient setting for a given environment.
Use Multiple Servers¶
OpenVPN is not multi-threaded so any single instance of OpenVPN is limited to using a single CPU. If a router has fast cores and not too many users, that may be OK, but it does not scale well. A workaround for this is to split users onto multiple servers. There are various means to reach this goal, including (but not limited to):
Multiple servers on different WANs or ports, each with unique tunnel networks but otherwise identical settings (Same CA structure, encryption, etc).
Administrators could choose to manually configure pools of clients to connect to specific servers, but that does not scale well.
Clients may connect to any server configured in this manner so long as their settings line up properly.
Multiple servers can be listed in a single client configuration with additional
remote-randomto the client configuration so that clients will pick a random server when starting, which avoids overloading whichever server is listed first.
Servers could be run on multiple WANs to overcome single-circuit throughput limits.
Multiple servers with completely unique settings (Different CA structure, different clients, etc)
More secure but more difficult to manage.
Clients must use different configurations to reach each server, no automated/built-in way to pick between them unless a specific client supports that function.
Good for isolating separate security levels (e.g. remote workers, remote administrators, vendors).
As a counterpoint to the above, each server will incur additional memory and other overhead to manage the process. When dealing with site-to-site VPNs, it is more efficient from a memory standpoint to use a single server with multiple clients (Peer to Peer SSL/TLS in client/server mode) vs servers for every node (Peer to Peer SSL/TLS with a /30 tunnel network). If memory is a limiting factor, use fewer servers. If CPU overhead is the limiting factor, use separate servers.