Load balancer positioning

[rajeshck]

We have a typical multi-site data center setup in which 2 X ECMP PLRs connected to the north Core networks and south UDLR. Workload/VM are spanning between 2 data centers and clients are located in WAN edge. High level, Data path as follows

For Targets in DC1

Client-->WAN-->DC1 Core-->either of DC1 PLRs --> UDLR -->Logical wire --> VM

For Targets in DC2

Client-->WAN-->DC2 Core-->either of DC2 PLRs --> UDLR -->Logical wire --> VM

In case or VMotion clients will follow the DC1 path but they will use DCI link for VXLAN overlays ( No ingress optimization )

Now we are in a position of placing a LB which can serve all purpose of Load balancing requests, means one EDGE will serve different VIPs for both one-arm and inline services. It would be great some design experts advise me the placement of the edge.

1. where exactly i shall place the edge? edge directly connecting to the core networks or connecting UDLR - my concern about the hair pinning or the data traffic
2. can the pools contains IP from different networks? consider a case of primary VM in DC1 and secondary VM in DC2 - what optimal design for such kind of requests
3. usual designs for LBs are close to the application, since these VMs are floating around what would be the best approch - please advise any reference design is available

Thanks in advance

Rajesh

[Sreec]

1. where exactly i shall place the edge? edge directly connecting to the core networks or connecting UDLR - my concern about the hair pinning or the data traffic

Unfortunately you don't have ingress optimization which is certainly a drawback if machines floats around and expectation is to have line rate performance . It is something you have to test it with current design and capture the results based on traffic pattern. Also if edges are in ECMP, LB is not possible.

1. can the pools contains IP from different networks? consider a case of primary VM in DC1 and secondary VM in DC2 - what optimal design for such kind of requests

         Haven't tested this, ideally clustered applications in L2 mode is preferred (This is not GTM solution) Do you have any heartbeat (L2)requirement for these servers ? 

1. usual designs for LBs are close to the application, since these VMs are floating around what would be the best approch - please advise any reference design is available

        One approach i can think about is having multiple pools - Active pool for Site A and Passive Pool for Site B . Failover to Site B pool when ever active Pool is not reachable. Or give maximum weight to Site A Machines , and less traffic to Site B.(whenever traffic  goes to Site-B  there might be a dip in the performance- so i repeat,test the configs carefully

[rajeshck]

Thank you Sree for the quick response, appreciate.

Is there any out-of-the-box way to do ingress optimization from VMWare?

1. I am still not clear the exact positioning of the LB EDGE, shall we directly hook into the core routing engine or create a separate logical wire and attach to the UDLR or an EDGE forked to ECMP PLRs.
2. There is no heart beat requirements - the reason i asked, i never seen any document or examples for the pool member VMs from different subnets, which obviously limiting the flexibility of subnets in distributed data centers. for example primary node in DC1 with 10.10.10.x and secondary node in DC2 with 10.20.10.x
3. I like the idea of adding nodes in different pools and using application rules to fail over - can you recommend good read on this?

Once again thank you for your efforts

Cheers!

[Sreec · Accepted solution]

1. I am still not clear the exact positioning of the LB EDGE, shall we directly hook into the core routing engine or create a separate logical wire and attach to the UDLR or an EDGE forked to ECMP PLRs.

          When you say, core routing engine - are you referring to DC core ? .  What type of users are going to access these load balanced machines ? MPLS or DC internal ?

1. There is no heart beat requirements - the reason i asked, i never seen any document or examples for the pool member VMs from different subnets, which obviously limiting the flexibility of subnets in distributed data centers. for example primary node in DC1 with 10.10.10.x and secondary node in DC2 with 10.20.10.x

If your apps are stateless, i don't find any issues . Happy to correct myself if i'm wrong .

1. I like the idea of adding nodes in different pools and using application rules to fail over - can you recommend good read on this?

Something like this ->Configuring NSX-v Load Balancer for use with vSphere Platform Services Controller (PSC) 6.0 - VMware...

[rajeshck]

When you say, core routing engine - are you referring to DC core ? .  What type of users are going to access these load balanced machines ? MPLS or DC internal ?

-- Yes, DC core - In our case, Cisco Nexus switches.

-- Requests from outside DC, MPLS and sometimes DMZ requests

Cheers!