I've created a 3-node Hyper-V cluster (2012 R2) out of 3 former standalone Hyper-V host servers, using a converged networking model andfollowing this setup almost exactly. I've added an iSCSI NAS (QNAP) set up as a CSV, and thefirst VM that I live migrated to cluster storage took 11+hours for about 1.5 GB of VHD files(!). There shouldn't have been any existing cluster-related traffic to even load-balance, and a straight copy of a file that size to the NAS should take < 20 sec. Briefly, here's my config:
- NAS has (4) teamed 1Gb NICs (in testing, I'm getting ~112 MB/s to a hosted file share).
- NAS is on our main internal network (10.10.10.x) because it will also be used for minor file sharing duties.
- Each cluster server has (4) 1Gb NICs. I've added 3 of them to an LBFO team without management OS access, and reserved the 4th for remote host access.
Question: Is this necessary? Unfortunately, when I enabled management OS access on the LBFO team adapter, each NIC on each server grabbed an IP via DHCP, and I want to avoid that. - Misc. acronyms: MPIO is set up and working per QNAP's docs, the servers/NICs support VMQ and RSS, but not RDMA or SR-IOV. I turned on Jumbo Frames across the fabric, but I ran into issues and the cluster seemed to break - turning it back off
seemed to fix things (I think it may be because QNAP supports a frame size of 9000 and the Broadcom NICs show 9014).
Question: I assume server restarts aren't needed - it's just a matter of turning on Jumbo Frames on the NAS, switches, and host NICs, correct?
TechEd webcasts and the dark depths of the Interwebz make me think my converged configuration is supported. With that said, my key questions are:
- Most importantly: It looks like MPIO & iSCSI is supported with LBFO-teamed NICs if you use the MS solution and not a vendor's. However, should I be trying to do afully converged solution that uses all 4 NICs, and for which there's a virtual Storage VLAN on the virtual switch for my NAS (like the MS best practices example, but with iSCSI instead of SMB, if that's possible)?Or should I dedicate 2 of the physical NICs to the NAS, and have the other 2 NICs in a team that handles the rest of the roles?
- Should I only have 1 MPIO session per 1 NIC in the team, or does that matter?
- I did add the "Management" virtual network in the sample configuration, but I just realized that's probably redundant with the dedicated "Host Access" NICs that I reserved. Should I remove that network, or leave it and add the extra NIC to the LBFO team?
- Are there additional QoS settings that I should check/tweak? Should I shift minimum bandwidth weightings just for the period that I'll be migrating VMs?
- Is there any way to get all of my clustered virtual networks imported into SCVMM 2012 R2 for management? The lowest level of entity that I can manage in a UI seems to be a virtual switch. That's fine, but if MS is encouraging adoption of converged networking, I would expect their tooling and UIs to have better support for it.
It seems as if the cluster is running fine in general, but the slow storage migration is killing me, and I assume it can only get slower as I add new VMs to cluster storage. My best guess is that the virtual networks aren't set up properly somehow...thoughts?