VLAN in the RB750Gr3 is about 310-350Mbit, and CPU load about 99% at one core (RB750Gr3).
Test is done in minimal setup, no fw roules. Test is with cat6 cable, directly from MT to two computers.
When testing on same bridge without VLAN, speed is about 750Mbit, and CPU load is low.
For now i don’t want ROS 7.x
Update: When doing more streams 3-8 are tested at VLAN, speed is about 700-750Mb between segments, so for me ok . And then it spread it over more CPU.
With RoS 6 the switch on hEX can’t do VLAN - it’s handled by the CPU, hence your problem. I think (THINK) that with RoS 7 the switch on the hEX would be able to handle this. But, for now, I wouldn’t recommend RoS for production.
RB4011 would brute force it, but I don’t think (not sure) that its switch will not do it in hardware with RoS 7.
RB5009 would brute force it, and I think that its switch chip would do this in hardware. But this device is RoS 7 only.
Traffic between VLANs is always routed. Some of the 3xx CRS (only cRs, not cSs) switches series will do routing in hardware - this means wirespeed, even if each port is a different network, But there are some limitation, and this will work only on RoS 7. Looks like the RB5009 switch will do this too. But only with RoS 7, and it is pretty much early stages - handle with care, and don’t blame me if it brakes later.
I don’t know. Finding a device that handles this is easy - the hard part is to find a cheap one.
Routing (also inter-VLAN) is L3 stuff, so L2 offload can not happen. Ability to delegate tagging and untagging for access ports to switch chip somehow improves throughput, but not that much, most work is still done by CPU shuffling packets between different VLANs.
Yes, routing is L3. But with RoS 6.x the L2 VLAN part of the hEX is CPU based - the switch chip isn’t used. So, he get L3 routing AND VLAN processing using CPU. Take out the VLAN processing, and the hEX should do 1gbps with fastpath. Mine does PPPoE + fastpath + firewall + NAT and goes over 800Mbps.
. And then it spread it over more CPU.