When doing live migrations, transfers do not break 1GB even though we have a 10GB connection from host to host. The transfer mostly hang around the 600 - 900 Mbps range. A normal VM takes 10 minutes or so to move. In the very last second or two of a migration, task manger will show a transfer speed around 9GB.
What can I do to get that kind of speed the whole time? Here is my setup and what I have tried.
(2) Windows 2012 R2 servers running only the Hyper-V roll.
Intel S2600GZ board.
(2) 8 core 2.6 GHz Xeon E5-2650 processors.
96 GB of memory.
(2) Intel X540-T2 10GB NICs. (4 10GB ports per server)
I've enabled jumbo packets and can successfully ping with an 8000 byte unfragmented packet from host to host.
I've disabled all C states in BIOS
These servers have a four port 1GB NIC in them. I've disabled that in BIOS.
I've checked the CPU during transfer - no core maxes out at all (doesn't bring 10% usage).
I've tried a variety of live migrations setups (with compression and without). No difference that I can see.
I've tried a direct connection between these two host with no switch in the center. No difference.
I've teamed the NICs in Windows - seemed to add about 300 MBs to the transfers (that is why I have the 600-900 range)
I've configured the NICs to use the Hyper-V profile in the Intel management software.
I've updated the firmware on the servers to the latest available.
Finally, I let SCVMM manage the NICs (knowing that I shouldn't expect more than 3.5 Gbps due to VMQ). The live migration traffic then gets spilt between three vNICs (live migration, host management, client access). I found this behavior very odd because I have the IP range used on the Live Migration network as the only specified range to use under Migration Settings.
So where should I start looking now?