On my MT7621 based router, I am using the latest 4.9.30 kernel with mini jumbo frame patch. It seems my ISP allows me to set MTU 1500 for the PPPoE connection.
At the moment I have eth0, eth0.1 (lan) and eth0.2 (wan) set to MTU 1508. On top of eth0.2 I get pppoe-wan which I can set to MTU 1500.
All my wlan-x connections remain MTU 1500
And the bridge interface I keep at 1500.
Is this correct?? And how should I notice the improvement that I can now speak standard MTU 1500 to the internet?? Besides the obvious check a PING -s 1472 whatever.com doesn't fail.
I assume you know, but as a service for other's stubling over this thread: for IPv4 you need to set the don't fragment bit as otherwise IPv4 fragmentation/defragmentation will hide the fact that your ICMP message exceeded the MTU...
C:\WINDOWS\system32>ping -f -l 1472 github.com
Pinging github.com [192.30.255.113] with 1472 bytes of data:
Reply from 192.30.255.113: bytes=1472 time=253ms TTL=47
Reply from 192.30.255.113: bytes=1472 time=252ms TTL=47
Reply from 192.30.255.113: bytes=1472 time=252ms TTL=47
Reply from 192.30.255.113: bytes=1472 time=242ms TTL=47
Ping statistics for 192.30.255.113:
Packets: Sent = 4, Received = 4, Lost = 0 (0% loss),
Approximate round trip times in milli-seconds:
Minimum = 242ms, Maximum = 253ms, Average = 249ms
C:\WINDOWS\system32>ping -f -l 1473 github.com
Pinging github.com [192.30.255.113] with 1473 bytes of data:
Packet needs to be fragmented but DF set.
Packet needs to be fragmented but DF set.
Packet needs to be fragmented but DF set.
Packet needs to be fragmented but DF set.
Ping statistics for 192.30.255.113:
Packets: Sent = 4, Received = 0, Lost = 4 (100% loss),
C:\WINDOWS\system32>
Note that this is the first time I (have to ) use an ISP with PPPoE. I was expecting to get more responsive webpages, faster downloads. An overall better experience. So I guess my setup is correct, just not what I hoped 
PPPoE with MTU of 1500 really ??
PPP frames are encapsulated into Ethernet frames.
If you expand these PPP to 1500 bytes the resulted frame is bigger as a standard ethernet frame.
The trick behind this is iptables, read here.
http://lartc.org/howto/lartc.cookbook.mtu-mss.html
Really, they are affectionally called baby-jumbo frames. Instead of going up to ~9000 bytes that "real"jumbo frames would allow, some ISPs allow up to 1508 bytes so that the PPPoE overhead does not eat into the MTU to normal internet sites (which typically are restricted to MTU = 1500). So one still does pay the 8 byte overhead out of the link bandwidth. but at least one can use MTU 1500 packets to other hosts on the internet.
Thanks for the hint.
Baby jumbo, I think this will be ~4000 or so.
Jumbo frames are upto 9000 bytes; the MT7621 can do max 2000 (2k ??) due a hardware limitation. For me it was interesting to see performance differences with the "standard" 1500 MTU inside the PPPoE.
I am experiencing slow loading websites depending on the website. I was hoping this would solve it. Seems like not 
I'm paying for 200/20 (max we can get here on fiber). On a good day Im getting even 250/22; on a bad day less then 50/20. That is with all else the same. I get 4K IPTV (well, the on demand movies, not live streaming). That runs on different VLAN and works well.
I will keep the MTU settings cause I feel it makes sense. Trace routes with MTU path discovery does show that it stays at 1500 to all the sites I tested.
So for now: back to the drawing board to figure out how to get good performance. (Probably a new topic when I experiment some more)