VMware Cloud Community
fduranti
Hot Shot
Hot Shot
Jump to solution

dvUplink/teaming and failover

I've a question related to how dvuplink work regarding link failover...

Let me say that I have this kind of configuration:

dvSwitch configured with 4 dvUplink (calling them dvu1 dvu2 dvu3 dvu4).

cluster 1 with 5 hosts need to use 2 of those dvuplink as Active/Active (and have 2 pnic configured on dvu1 and dvu2)

cluster 2 with 5 hosts need to use the other 2 of those dvuplink in Active/Standby (dvu3 is a 10g and the standby adapter dvu4 is only 1g)

the portgroup with vkernel interface is configured so that it has:

Active dvuplinks:

dvu1

dvu2

dvu3

Standby dvuplinks:

dvu4

On the second cluster (the one with 10g and 1g) it seems that I see traffic on both cards.

I was expecting that having only dvu3 and dvu4 configured the dvu4 will not go IN as a failover for dvu1 or dvu2 because those uplink are not even configured on that host.

Configuring only 2 dvuplink port and putting them in active/standby work as expected.

Is this the normal behaviour of the dvUplink ports?

Reply
0 Kudos
1 Solution

Accepted Solutions
lambeth
Hot Shot
Hot Shot
Jump to solution

Yes that is the way the teaming algorithm in ESX works. Your policy is basically interpreted as "try to keep 3 active uplinks". ESX will bring in a standby anytime there are not 3 active uplinks, and hosts with no pNIC connected to an uplink port get treated the same as if there were a pNIC there with link state down. The only way to accomplish what you really want currently would be to use separate vDS for the separate host types.

One question for you though, presumably you're trying to configure this way bc you want ESX to favor the 10G uplink in the hosts which have that. Would introduction of a IO scheduling which considered uplink speed so that for example if you teamed a 1G and 10G uplink active/active you'd see a 10:1 ratio of traffic favoring the 10G meet your needs? Or do you envision still having a need to specify teaming on a per host basis rather than per portgroup?

View solution in original post

Reply
0 Kudos
2 Replies
lambeth
Hot Shot
Hot Shot
Jump to solution

Yes that is the way the teaming algorithm in ESX works. Your policy is basically interpreted as "try to keep 3 active uplinks". ESX will bring in a standby anytime there are not 3 active uplinks, and hosts with no pNIC connected to an uplink port get treated the same as if there were a pNIC there with link state down. The only way to accomplish what you really want currently would be to use separate vDS for the separate host types.

One question for you though, presumably you're trying to configure this way bc you want ESX to favor the 10G uplink in the hosts which have that. Would introduction of a IO scheduling which considered uplink speed so that for example if you teamed a 1G and 10G uplink active/active you'd see a 10:1 ratio of traffic favoring the 10G meet your needs? Or do you envision still having a need to specify teaming on a per host basis rather than per portgroup?

Reply
0 Kudos
fduranti
Hot Shot
Hot Shot
Jump to solution

Thanks lambeth for your answer Smiley Happy

My problem is exactly the one you explained. I've 2 cluster that, one with 2 1GB nic and one with 1 10GB and 1 GB nic configured for the vmkernel and some vmnetwork. We're using vmkernel not only for vmotion but also to mount all of our NFS datastore and some VM have an interface on portgroup on that dvSwitch to use them with NFS mount too.

For us the best solution for us was to configure 1 dvSwitch with our different VLAN and making them to work in the better possible way.

One of the problem is also that the cluster that use 2 1GB nic have those nics configured in ip load balancing with etherchannel on 2 cisco 3750 stack switch. This is also creating us a problem because be cannot even configure the portgroup with "port lb" or just failover because from what we saw if you change the Load Balancing away from ip balancing the traffic get strange and in some cases we also loose visibility of datastore (this was our old configuration with etherchannel also on the storage).

At this moment we've configured 2 different portgroup for each VLAN on our cluster so that one use 2 interface in ip-balancing and the second use the interface in an active standby configuration.

Probably we will change the configuration on the first cluster to virtualport id balancing or active/standby on this weekend to unify the vlan.

Your question about the introduction of an "IO scheduling" that take in consideration the NIC speed is really interesting. I think it could be a good solution so that we can do a unique configuration with all kind of nics, in the next few weeks we will get some servers with 2 10G network card so we will have 3 kind of network connections on our hosts. But probably the better will be to have the choise from IO scheduling and per host portgroup.

IO scheduling will be simpler to configure just do a virtualport id balancing and the traffic will flow in a "good" way.

The possibility to configure per host configuration instead can allow better performance/fine tune of configurations like:

1) some hosts with 2 10g in virtual port id balancing or with specific configuration (vmotion/ftlog on 1st 10g, iscsi/nfs on 2nd 10g with each card that is the standby of the other one)

2) some hosts with 10g/1g in active/standby for all vlan

3) some hosts with 2 (or more) 1gb on a single switch in ip-load balancing/virtual port id balancing or whatever configuration is appropriate for them

Reply
0 Kudos