Why the copy transfer speed between instances is different when only a single instance see the other?

[Luken (Łukasz Zaroda)]

There is this old comment mentioning this on lxd:

Slow copy/move between lxd hosts

I apologise for necroposting, but maybe it may be helpful for someone in future. The advice is simple: make sure both nodes can connect to each other, not just the one you’re transferring the container from to a new one. Once I enabled a reverse connection in the source machine’s firewall and added the source machine as a remote for the destination one, the transfer speed skyrocketed for some reason.

But I just experienced it on Incus. I was copying containers in a push mode because the source instance had access to the incus port of the target instance, but not the other way around. It was like 6MB/s. Then, on a source instance, I added a firewall rule allowing the target instance to “see” the incus port on the source, and the speed jumped to like 40MB/s for the next container. What sorcery is this? Is this expected behavior?

incus 6.0.2
zfs to zfs

[stgraber (Stéphane Graber)]

Likely something affecting TCP window size scaling in one direction.