Hi Stuart, > > As seen in dmesg (see below), both servers are guest machines, and as > > such use vio(4) (more precisely a vlan(4) on top of vio(4)). > > > > We did notice in the 7.6 changelog that some changes have been made to > > vio(4), but are unsure if they come at play here. > > Yes. > > Looks like pascal@ figured this out. Another area where network offload > features are broken. > > Try "sysctl net.inet.tcp.tso=0".
Thanks a lot for the pointer, but disabling TSO only was inconclusive. As the 7.6 changelog also mentioned enabling LRO for vio(4), we did give it a try by disabling it on vio0: ifconfig vio0 -tcplro And performances came immediately back to be on-par with the 7.5 IKEv2 server. We ran the same tests again with TSO enabled, without any noticeable difference, so it looks like LRO was the one at fault here. So problem "mitigated" successfully here. Is there any other test which may be worthy to conduct here? Loïc REVEST
