4 ms·

FYI WireGuard is pretty slow on the Edgerouter compared to IPSec. I assume IPSec has a hardware offload and WireGuard does not.

JustSomeNobody 7y ago
Yes, IPSec is supported for HW offload: https://help.ubnt.com/hc/en-us/articles/115006567467-EdgeRouter-Hardware-Offloading https://help.ubnt.com/hc/en-us/articles/115006567467-EdgeRou... Disappointed to hear the performance hit with WireGuard. I was planning on trying it out, but I think I'll wait.
- zrm 7y ago
  Note that hardware offload only really matters when you're combining fast networks with slow processors. WireGuard on a modern PC CPU will more than saturate gigabit ethernet and the per-packet latency hit is <1ms.
- msh 7y ago
  Well for some of their routers its false advertising. On the edgerouter x ipsec offload for client vpn have been broken for more than a year.
elabajaba 7y ago
According to https://an.undulating.space/post/181227-er_alternate_firmware_vpn_benchmarks/ https://an.undulating.space/post/181227-er_alternate_firmwar... Wireguard is just as fast as hardware accelerated IPSec on both the Edgerouter X and Lite. With Openwrt on the Edgerouter Lite Wireguard is ~2x faster than hardware accelerated IPSec on EdgeOS.
- wahern 7y ago
  They don't give the IPSec parameters. Wireguard uses ChaCha20 for encryption with Poly1305 for authentication. IPSec can use many different combinations of cipher and authentication algorithms. If they were using AES+SHA256 then SHA256 would likely be the bottleneck. AFAIU, Ubiquiti routers use a crypto coprocessor (as opposed to AES-NI and SHA-NI on amd64), which means even if SHA256 was accelerated on the coprocessor it could easily still be too costly. But we don't even know if the MAC was accelerated at all. The implication that "IPSec acceleration" handicaps Wireguard is unfounded. If those benchmarks were with AES+GCM, then that would definitely be surprising. But the safe bet is that they were using AES+SHA256 (or something other than GCM for the MAC), in which case their benchmarks are not surprising, simply misleading. IPSec can also do ChaCha20+Poly1305 (at least, OpenBSD's stack can). Any serious comparison should also include IPSec using the same crypto algorithms as Wireguard.
- covercash 7y ago
  I’d love to see benchmarks on an ER-4/6/8, especially with gigabit WAN availability becoming more common.

JustSomeNobody 7y ago

Yes, IPSec is supported for HW offload: https://help.ubnt.com/hc/en-us/articles/115006567467-EdgeRouter-Hardware-Offloading https://help.ubnt.com/hc/en-us/articles/115006567467-EdgeRou... Disappointed to hear the performance hit with WireGuard. I was planning on trying it out, but I think I'll wait.

zrm 7y ago
Note that hardware offload only really matters when you're combining fast networks with slow processors. WireGuard on a modern PC CPU will more than saturate gigabit ethernet and the per-packet latency hit is <1ms.
msh 7y ago
Well for some of their routers its false advertising. On the edgerouter x ipsec offload for client vpn have been broken for more than a year.

elabajaba 7y ago

According to https://an.undulating.space/post/181227-er_alternate_firmware_vpn_benchmarks/ https://an.undulating.space/post/181227-er_alternate_firmwar... Wireguard is just as fast as hardware accelerated IPSec on both the Edgerouter X and Lite. With Openwrt on the Edgerouter Lite Wireguard is ~2x faster than hardware accelerated IPSec on EdgeOS.

wahern 7y ago
They don't give the IPSec parameters. Wireguard uses ChaCha20 for encryption with Poly1305 for authentication. IPSec can use many different combinations of cipher and authentication algorithms. If they were using AES+SHA256 then SHA256 would likely be the bottleneck. AFAIU, Ubiquiti routers use a crypto coprocessor (as opposed to AES-NI and SHA-NI on amd64), which means even if SHA256 was accelerated on the coprocessor it could easily still be too costly. But we don't even know if the MAC was accelerated at all. The implication that "IPSec acceleration" handicaps Wireguard is unfounded. If those benchmarks were with AES+GCM, then that would definitely be surprising. But the safe bet is that they were using AES+SHA256 (or something other than GCM for the MAC), in which case their benchmarks are not surprising, simply misleading. IPSec can also do ChaCha20+Poly1305 (at least, OpenBSD's stack can). Any serious comparison should also include IPSec using the same crypto algorithms as Wireguard.
covercash 7y ago
I’d love to see benchmarks on an ER-4/6/8, especially with gigabit WAN availability becoming more common.