Network Virtualization using Extreme Fabric Connect

Network Virtualization Using Extreme Fabric Connect
© 2019 Extreme Networks, Inc. All rights reserved. 117
Distributed Virtual Routing
This section will go into greater depth about Distributed Virtual Routing (DVR) in the context of the
Extreme Fabric Connect architecture. DVR is an enhancement of Fabric Connect that allows both the use of
a distributed anycast gateway and the ability to compute the shortest path to the individual host IP. This
becomes hugely important in environments where the host IPs are mobile. The data center is the most
challenging environment in this respect since the advent of server virtualization. VMs can be moved with
ease, without even being stopped, from one physical hypervisor to the next and indeed across geo-
redundant data centers. When those VMs migrate, they always take their IP address with them; this is
necessary lest the applications running on those VMs would lose all their connections (open sockets) in the
process.
Traffic Tromboning Challenges
The best way to understand the benefits of DVR is to first of all understand what would be the limitations of
a Data Canter Fabric Connect architecture if it was not DVR enabled.
We already know that the SPB Fabric by definition always calculates the shortest path. But it will always
calculate the shortest path towards some already defined Backbone MAC (BMAC) which constitutes the
destination BEB for the traffic at hand. But since all traffic carried over the SPB Fabric is part of a service
type (VSN), and the addressing used in these virtual networks is not the BMAC but some other L2 or L3
addressing scheme, there has to be a mapping between the two.
Hence if we look at L2 VSNs, these services learn end-user MAC addresses (CMAC) and associate these
with the BEB’s BMAC from which they have been learned. The assumption is that a given CMAC is behind a
given BMAC and thus taking the shortest path to that BMAC also ensures the shortest path to that CMAC.
This works fine for L2 flows that remain within the same L2 segment (source and destination are in same IP
subnet) and is equally applicable to L2 flows within the data center.
Similarly with L3 VSNs, they exchange IP routes via IS-IS and install these IP routes in the relevant VRF IP
routing table. In this case, it is an IP network that is being associated with the BEB’s BMAC that has a local
IP interface on that IP network. Again, the assumption is that any host with an IP address belonging to that
network is residing behind that BEB and that taking the shortest path to that BMAC also ensures the
shortest path to that IP address. This assumption generally holds true in the campus, but does not hold true
in the data center where VM hosts are highly mobile in the east-west direction along server VLAN L2 VSNs.
Figure 64 illustrates the worst-case scenario of what could happen in an architecture where two geo-
redundant data centers have been made into a single SPB Fabric, leveraging Fabric Extend, and where DVR
is not being used. The server VLAN is L2 extended across both data centers using an L2 VSN. Both core
routers in both data centers have an IP interface on the server L2 segment and act as default gateways for
that same segment. They thus all announce the server subnet northwards toward the wider campus and
the branch offices in the example. They will also have to use some form of gateway redundancy protocol,
like standard VRRP, southwards toward the data center hosts (VMs). This can result in two forms of traffic
tromboning.
In the first case, traffic from a branch office destined to a data center VM will make a forwarding decision
based on the lowest cost to reach one of the four IP routers announcing the VM’s IP network, but since
there is no knowledge about where the VM is actually located, there is a 50% chance that the traffic will
arrive to the wrong data center and then need to take a second high latency hop over potentially the same
WAN (if the L2 VSN is Fabric Extended between the two data centers) to reach its destination.
In the second case, any data center L3 flow that needs to be IP routed either to reach another host in the
data centers (L3 east-west flow) or to reach the wider campus (south-north) will need to hit the default
gateway for the server segment. With standard VRRP, only one IP router will be acting as that default