Customer's workplace has no Internet provider, so he has installed a small custom all-Mikrotik network to feed Internet from his home, approx. 3km away using a pair of SXT5sq in transparent link mode (b/c), a POE switch in the shop (d), and a small MT router (in bridge mode) to supply WiFi signal (e). The entire network was bridged to his home router (a), which is a Netgear, the only non-MT unit. He ran bridged so he could access all his devices from either location, just like one might to a shed or other outbuilding. All networking units were hardcoded with static IP addresses in 192.168.1.0/24 outside the DHCP range configured in the Netgear. Non-networking units in the shop fetch IP addresses from the home Netgear.
Some time after establishing this link, an envious neighbor in the work complex asked him to share a taste of his Internet; so he set up a second link (f/g) between their two shops, about 1km, still bridged, same addressing rules. Inside the neighbor's shop is a SOHO router, double-NATting, but no one cares. So now the network looks like this:
Traffic works fine, but there is a weird speed issue.
The network will push about 30-45Mbps (which is plenty, as it exceeds his home gateway speed) in each direction over almost any combination of links. The exception is end-to-end: unit b to unit g, which runs spastically around 10Mbps. If speed is measured from c to g, or b to f, full speed is seen. The only time the speed drops is when both wireless links are in play.
I've assured myself that both links are on different channels, so it's not an interference issue. I have turned off STP/RSTP on every device in the network (save possibly the MT POE switch, which has a global RSTP setting that I can't puzzle out, and refuses to upgrade from 2.4). I've assured there are no duplicate IP addresses anywhere in the network. It's not the CPU load of bandwidth test, as I can get fine speed results between any two SXTs other than the two farthest apart. I've disabled g's Ethernet link and run the speed test, just to ensure that nothing going on inside the neighbor's shop is doing something strange, and results are the same. There are only about a half dozen devices in the main shop, so it's not a bridging overload. The wireless distances involved here are quite modest.
Why should the speed drop precipitously only in the end-to-end case? I feel like i'm missing something basic, but I can't find it.