Implementation of LACP in vDS without loss of network connectivity

[ABusch]

Hello,

Situation: a 2 node esx Cluster with one vDS with 4 physical uplinks. The vDS has 4 portgroups (management, 2 for vmotion
and one PG for vm network).

What would be best practice to implement LACP without interrupt of network connectivity?

I have read about some problems removing some uplinks from an existing vDS. So anybody done this?

Best

Alex

[chriswahl]

What is your use case for implementing link aggregation? Do you only wish to use it for the virtual machines?

Any time you remove an uplink there is a potential for network disruption. This stems from the upstream switches having to update their MAC address tables after the host sends a RARP. Typically it's less than a ping, but you might lose a ping (or two) as the physical infrastructure learns the MACs.

[ABusch]

Dear Chris,

the use case should be more bandwidth for the network of the vms. I think the isolation of the vmotion traffic (not only through VLAN) through binding to a dedicated nic is not a bad idea. So at least there should be 2 uplinks which are in an LACP aggregation.

I think a loss of a packet is not a problem because the removal of an uplink should not occure every day.

Thanks

Alex

[chriswahl]

the use case should be more bandwidth for the network of the vms.

Even with LACP, each IP session is still limited to a single uplink. Unless your VMs are initiating sessions with multiple unique IP addresses simultaneously and are currently hitting the throughput limitation on a single uplink, you won't realize any benefit from using a LAG. One alternative is the use of "route based on physical NIC load" which will actively analyze uplink bandwidth saturation and migrate VMs across the Active uplinks when 75% traffic saturation occurs for greater than 30 seconds.

Also, if you plan to move forward with a LAG, you'll need to ensure that all uplinks on the VDS are enabled for the LAG. Move the non-LAG uplinks to a different VDS. Mixing uplink types in a VDS is not supported.

[ABusch]

So you wouldn't recommend using LACP in this case, ok. The alternative with the route based on NIC load I have already implemented.

But what should be the exact order of steps for implementaion and is this possible without downtime. Maybe you've done this or know any best practice. I mean (dynamic) LACP was introduced in esx 5.1 (if i'm right) and I think some people may have made an implementation on existing systems.

1. Removing 1 or 2 uplinks from the existing vDS (is this without interruption)?
2. creating a new vDS with these free uplinks and LACP.
3. migrating the portgroups/networks

[chriswahl · Accepted solution]

Here is a video that I created on using LACP with the VDS 5.1.

Configuring LACP on a vSphere 5.1 Distributed Switch - YouTube

Like I said prior, any time you remove an uplink there is a potential for network disruption. You would then need to create new distributed port groups that are required.

High level:

1. Create the new VDS
2. Remove 2 uplinks from the old VDS (chance for ping drop)
3. Add the 2 uplinks to the new VDS
4. Configure LACP on the physical ports and new VDS
5. Create needed port groups on the new VDS
6. Migrate VMs (chance for ping drop)