Sheepdog and dedicated nic

[webfrank]

Hi, Im experimenting with sheepdog and I would like to know if is it possible to redirect sheepdog traffic over dedicated nic. Sheepdog relies on corosync and I think all cluster traffic should go on same nic, or not?

 

[dietmar]

Would you mid to ask such questions on the sheepdog mailing list?

 

[webfrank]

Question has been asked on sheepdog ml and answer is to modify corosync configuration to have totem listening on desired ip address/interface but on proxmox this is not possible as cluster configuration in managed differently. Also changing cluster config I could not find a way to have corosync heartbeat on a nic and sheepdog replication on another one.

 

[dswartz]

I ran into this same quandary while playing with sheepdog the last couple of days. I had a thought on this: my motivation is to remove sheepdog replication traffic from the nics the guests on my 2-node proxmox cluster use for their network traffic. So, given that, instead of a dedicated link for that purpose, have a dedicated link for corosync (and therefore sheepdog). It would have to be visible to my workstation, but I can live with that. So something like this: nic0 => subnet 10.0.1.0/24 for corosync/sheepdog and nic1 => subnet 10.0.0.0/24 (current subnet) for guest traffic.

 

[mo_]

Im not sure how sheepdog works, but the "normal linux" approach should work since proxmox is a normal linux installation after all. What I mean by that is the following:

I assume you use (a) dedicated IP address/es to access the sheepdog cluster, yes? If so you can give the interface you want the traffic to go over an IP address in the same subnet/vlan and use the linux routing table to ensure that traffic for the sheepdog address/es uses said interface (by giving that route the lowest metric)

 

[dswartz]

sheepdog seems to just use the same interface as corosync. So I am wondering if I can change *that* address (well subnet) to use the dedicated interface. My main concern is I still need to be able to get to the management interface, but it looks like that is listening on the wildcard address, so that might 'just work'. I'll post my findings later...