Here you can see my situation:
It's a simple implementation of Q in Q with a c-vlan inserted into a s-vlan realized by bridges in a hyperlan transport network.
Routers/bridges are mikrotik RB750UP and RB450G and radios are ubiquiti stuff with L2MTU set to 2024.
This network works: you can ping from c-vlan IP1 to c-vlan IP2 with 1500 bytes packets without fragmentation.
The problem is latency:
Average of 20 1500bytes pings results (don't fragment):
1) from Hyperlan IP1 to Hyperlan IP2: 9ms ;
2) from S-VLAN IP1 to S-VLAN IP2: 12ms ;
3) from C-VLAN IP1 to C-VLAN IP2: 35ms ;
Average of 20 30000bytes pings results :
1) from Hyperlan IP1 to Hyperlan IP2: 170ms ;
2) from S-VLAN IP1 to S-VLAN IP2: 117ms ;
3) from C-VLAN IP1 to C-VLAN IP2: 117ms ;
As you can see, with 30k icmp packets (fragmented in 20 1500B, of course) you have the same latency in c-vlan and s-vlan, but more latency in direct eth IP pings;
With 1500B icmp packets the situation is much more different: s-vlan latency similar to direct eth one, c-vlan latency about 3x slow!
Why?
Using packet sniffer on eth interface you have always 1522 bytes packets (1508 + 14 eth), as with single 1500B pings that with 30k pings.
.