Oha, that sounds weird, could you describe the 860B1's port layout a bit more in detail. If eth0 is the only true interface from the SoC to the switch and it uses two VLANs to attach WAN and LAN to the switch, the instantiating sqm on the physical interface below those two VLAN interfaces is certainly the wrong thing to do. In my hypothetical? case both ingress and egress would be shaped to the minimum of your defined bandwidth as WAN egress traffic is LAN ingress traffic, plus for concurrent traffic things will be even worse. I hope that eth0.1 and eth0.2 are just weird ways to name two independent interfaces. Please feel free to elaborate about the specifics of your router. This is pretty important to get right, but also impossible without user input...
This shows that your system encounters giant (meta-packets), cake should be able to deal with those, but you might want to disable TSO, GSO and GRO (on all interfaces) to test whether these cause issues. See https://forum.openwrt.org/t/issue-with-ipv6-6in4-and-sqm/1348 for some more information about GRO disabling...
For DOCSIS/cable links the current recommendation is to use 18 bytes for the overhead (as that is the overhead that the DOCSIS mandated user-traffic shaper accounts for, the fact that DOCSIS adds more overhead is irrelevant for sqm-sxripts configuration).
You might want to have a look at the current version of https://lede-project.org/docs/howto/sqm, espevially the "sing and dance" section at the end on how to get the most out of cake (which admittedly is less urgent than getting it to work in the first place 
)
Best Regards