In my point of view, this would be better solved using by the concept that uses Juniper.
Al the services in a different namespace by default.
Some kind of linking-bridge between that namespace and the rest of the world.
In that case, if the services were in vrf=management, to other services reach that service would be obligatory some kind of route-leaking between those vrfs.
What is expected in a true scenario of VRFs.
When you talk about PPPoE in single thread, you are talking about what case specifically?
PPPoE-Client?
PPPoE-Server with only one customer?
I’m quite sure that PPPoE Server load can be distributed over several threads…
Even on RouterOSv6 it is possible. On older releases you could need some ugly thing as put several listeners on same interface, but it is possible.
But a single PPPoE connection (server or client) cannot be distributed on several threads.
And, maybe I’m mistaken, but it is valid to other PPPoE engines like Accell.
This is relative to the machine-state of PPPoE.
The solution to this is not multithread the PPPoE!
The solution to that, in the specific case of CCR2XXX that has Switch Chip, is Hardware Off-Load the PPPoE forwarding.
Your request should change a bit!
Instead of requesting Multithread, request hardware offload to then!
MPLS For example…
CRS can do Label Swapping (P. Router position) in line rate using the hardware offload and a very small CPU!
But the MPLS encapsulation (PE position) even CCR 2116 or a X86/X64 can do it in Line Rate without gaps for large amount of bandwidth and PPS.
So, your efforts with MikroTik team should be directed to convince then to use correctly the SDK of the Marvell Switch Chips.
Saying that considering white papers of Marvell says that those swit chips supports MPLS, and that there are other vendors in market that using that same switch chips can do hardware-based forwarding of MPLS.
An important thin note about VRF and Hardware Off-Load
Another thing that needs to be said is that VRF and Hardware Off-Load did not work well together on RouterOS.
A few months ago, I did some Prove of Concepts using VRF on CCR2004 and RouterOSv7.x(In my mind I remember being 7.12 or similar).
When I enabled VRF, it disabled the hardware offload.
Docs of Mikrotik about VRF confirms that, at least by now.
“VRF N/A Only the main routing table gets offloaded.”
Hey? What? Only on Main Table?
What about the concept of Underlay and Overlay?
Just to confirm…
This happens with a Vlan as a Sub-Interface of a hardware interface, or in a Bridge Vlan scenario?
In my opinion(guess) this is related to the fact that RouterOS allows you to do a fruit salad on L2 and L3 sub-interfaces.
On a full CPU scenario it does not affect much, but in hardware off-load this is problematic.
No other vendor with high hardware performance allows that mix.
What some calls of “magic and flexibility” is what causes the issues.
I think it is appropriate to require multithreading for these protocols. Mikrotik currently sells a lot of multicore devices that does not have the ability to support L3 hardware offload.
Thank you for sharing your experience, disabling one QSFP at a time results in the problem continuing (tested with only QSFP1-1 and QSFP2-1. I’ll try without the peer link being LACP and alternatively with a non bonded SFP+ interface as the peer link.
I have a suspicion that the issue is unique to the CRS354, sometimes works for hours without issues but can be as bad as loosing the peer link 20 times an hour.
Absolutely need software based offloads as well. Whether its Multicore FastPath or some new method, it needs to be done.
A lot of WISP’s are using RB4011, RB5009, CCR2004-16G-2S+ in their POP’s and would get no benefit from ASIC based offloads, they would however get a huge benefit from better software based offload support.
What is that even supposed to be, “software based offload”?
What we have now is a situation where PPPoE client is handled in software. The ethernet frames are received just like IP, and then the PPPoE header is stripped off in software.
It looks like this causes some load on the CPU limiting the datarate to below 1Gbps on many routers.
Other manufacturers use the same SoC in their routers and do not have that problem, so apparently there is capability in the SoC to offload part of this work to hardware. But MikroTik does not use it because they use standard Linux network handling instead of the dedicated SDK from each SoC manufacturer (understandable because of the many different SoC used in different MikroTik routers).
So now we have to hope that at some point, for the router models that support it and that are most affected by this, this offloading is supported by RouterOS.
But it would still be hardware offloading.
All those manufacturers that make the routers that ISPs supply with the 1Gbps+ internet connections they deliver today.
AVM, Arcadyan, Linksys, etc etc etc.