In my previous article vCloud Director with NSX: Edge Cluster I described various design options of NSX Edge Cluster in vCloud Director environment. In this article I would like to discuss additional option which extends the Design Option III – Dedicated Edge Cluster. Below is the picture showing the scenario from the previous post.
There is one Provider deployed Edge in the Edge Cluster for each Transit vCloud Director External network to which Org VDC Edge Gateways are connected to. The option works quite well for use cases where the Provider Edge is dedicated to single tenant – e.g. it is providing VPN services or L2 bridging. (Note that in L2 bridging use case the Org VDC Edge Gateway is not deployed and Org VDC networks connect directly to tenant dedicated external network).
However when we want to provide access to a shared service (for example internet) where we will deploy multiple Org VDC Edge Gateways of different tenants connected to the same external network they will all have to go through a single Provider Edge which can become a bottleneck.
As of NSX version 6.1 Edge Gateways can however be deployed in ECMP (Equal Cost Multi-Path) configuration where we can aggregate bandwidth of up to 8 Edges (8x10GB = 80 GB througput). High availability of ECMP Edges is then achieved with dynamic routing protocol (BGP or OSPF) with aggressive timing for short failover times (3 seconds) which will quickly remove failed path from the routing tables.
The problem is that (as of vCloud Director 5.6) Organization VDC Edges are deployed in the legacy (vShield/vCNS) mode and do not support ECMP routing nor dynamic routing protocols. The design I propose will get around this limitation by deploying Distributed Logical Router between Provider and Organization VDC Edges.
The picture above shows two Provider ECMP Edges (can scale up to 8) with two physical VLAN connections each to upstream physical router and one internal interface to the Transit Edge logical switch. Distributed Logical Router (DLR) then connects the Transit Edge logical switch with the Transit vCloud Director External Network to which all tenant Org VDC Edge Gateways are connected to. The DLR has ECMP routing enabled as well as OSPF or BGP dynamic routing peering with the Provider Edges. The DLR will provide two (or more) equal paths to upstream Provider Edges and will choose one based on hashing algorithm of source and destination IP of the routed packet.
The two shown Org VDC Edge Gateways (which can belong to two different tenants) then will take advantage of all the bandwidth provided by the Edge Cluster (indicated with the orange arrows).
The picture also depicts the DLR Control VM. This is the protocol endpoint which peers with Provider Edges and learns and announces routes. These are then distributed to ESXi host vmkernel routing process by the NSX Controller Cluster (not shown in the picture). The failure of DLR Control VM has impact on routing information learned via OSPF/BGP protocol even if DLR is highly available in active standby configuration due to the protocol aggressive timers (DLR control VM failover takes more than 3 seconds). Therefore we will create static route on all ECMP Provider Edges for the Transit vCloud Director External network subnet. That is enough for north – south routing as Org VDC subnets are always NATed by the tenant Org VDC Edge Gateway. South – north routing is static as the Org VDC Edge Gateways are configured with default gateway defined in the External Network properties.
The other consideration is placement of DLR Control VM. If it fails together with one of ECMP Provider Edges the ESXi host vmkernel routes are not updated until DLR Control VM functionality fails over to the passive instance and meanwhile route to the dead Provider Edge is black holing traffic. If we have enough hosts in the Edge Cluster we should deploy DLR Control VMs with anti-affinity to all ECMP Edges. Most likely we will not have enough hosts therefore we would deployed DLR Control VMs to one of the compute clusters. The VMs are very small (512 MB, 1 vCPU) therefore the cluster capacity impact is negligible