none
2012 SP1 Redirected CSV bandwidth reduction problem RRS feed

  • Question

  • I have a 3 node 2012 hyper-v cluster. It has 2 network interfaces set to 'use for cluster + client can communicate' and one dedicated for 'use for cluster' with the lowest metric (the preferred CSV network). All network connections are 1Gb, all teamed with at least 2 ports each.

    DPM backs up to tape only and is on a dedicated physical server.

    When the disks are owned by a single node and all VM's are also on that node, the 'production LAN' is saturated and the backup gets the best possible speed, clearly showing the network is the limiting factor.

    When some VM's are on another node, but the disks still owned by 'node 1', they are routed via node 1 during the backup process via the CSV network from the other node. I expect this to be the case. The strange thing in this case is that the bandwidth is measured at roughly 33% utilisation against the CSV network. All other interfaces remain idle.


    If I take one of the other cluster networks which has 'use for cluster + client can communicate' and remove it from cluster usage, the CSV bandwidth increases to 50%. 

    Clearly, the amount of nics that are designated to participate in cluster communication seems to then have a bearing on how much bandwidth the CSV network uses. If I had 4 I would expect 25% usage in the current scenario.

    I thought this may be to do with SMB multi channel support in 2012, as I know that when doing a basic file copy from node 1 to node 2, the bandwidth is split across all 3 network interfaces, roughly in 33% chunks. I disabled SMB client and server multichannel on all Hyper-V nodes and when doing a file copy from a specific network from node 1 to node 2, only that interface was used - as per typical behaviour pre 2012.

    However, it made no difference to the CSV bandwidth behaviour when VM and disk owner node differ - still at 33%. 

    Friday, October 3, 2014 3:05 PM

All replies