L2TP/LACP over multiple PPPoE links

RouterOS
1

Hello,

for a project i’m thinking about combining multiple PPPoE-links (VDSL) to a remote site for aggregate performance.

I was thinking about using 2 CCR2116-12G-4S+ (one at each end).

The remote site is house in a datacenter, connected via 2x25G to the internet, doing BGP to announce A /24 block and do some regular routing for the systems there.

On the other end i want to combine 5-10 VDSL-links (100/40) by starting a L2TP-session over each of the links, combining them via LACP with L3/L4-transmit hash.

EDIT: Or is Roundrobin better suited? Does it detect downed L2TP-links and skips them automatically?

Will that work or are there better solutions?

I want to achieve a nearly dynamic link that is immune to some of the PPPoE-links dropping out and is able to get more links aggregated as the demand grows.

Kind regards

Stephan

2

ECMP is much better in many aspect then playing with MLPPP. Simply bring up all of the links and use ECMP over them. the router statistically loadbalancing over them if you have enough flows. This kind of LB technique is independent from underlayer technologies. You can combine a PPPoE and L2TP link, or any other tunnels.

3

Am i able to use the cumulated bandwidth of all links with a single tcp connection or only for multiple connections?

This is one thing that would be really nice.

4

One TCP/UDP/ICMP/etc connection is one flow. For ECMP, you must have multiple flows. Per-packet load-balaning, however possible, it messing up some things, especially in TCP, so to be avoided.

5

Ok, but what’s the issue with multiple links, same rtt to the target and round-robin?

6

For example packets out-of-order.

7

Isn't this an inherit problem of the Internet and should be covered by the IP stack?

8

If this were the case, not every manufacturer and engineer would avoid this type of operation. This is not a problem on a non-critical low-speed links (<100Mbps), as this can handled only in CPU and increasing latency and jitter much. OoO handling in ASIC is very expensive. OoO packets simply drops in HW.

9

But what has the asic to do with it? This should be handled by the IP stack of the operation system behind the router. We are not talking about fragmentation!