2

On a Dell PE T610 I have Hyper-V core running, with 2 x Broadcom BCM5709C NetXtreme II GigE installed. I have noticed that copying large files 17GB for example, from a network physical server to the Hyper-V host local drive [not vm guest] is very slow in comparison to copying from Physical to Physical servers.

  • Copying a 17GB file physical to Hyper-V host takes 30 minutes
  • Copying a 17GB file physical to physical host takes 15 minutes

Can someone tell me exactly what registry nodes I should disable on Hyper-V NICs to improve performance. So far I have gone to HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Control\Class\{4 D36E972-E325-11CE-BFC1-08002BE10318} and set the following to 0 on both physical NICs:

  • LSOv1IPv4
  • LSOv2IPv6
  • TCPUDPChecksumOffloadIPv4
  • TCPUDPChecksumOffloadIPv6

Should I also disable *TCPConnectionOffloadIPv4 & *TCPConnectionOffloadIPv6?

Starfish
  • 2,716
  • 24
  • 28
gary
  • 191
  • 1
  • 8
  • Do you have the latest drivers on there? I've got several Hyper-V boxes running the 5709 NICs (teamed) and they're blazing fast. Also, is your storage subsystem up to the task (I only ask because you didn't mention anything about it)? – Chris S Jul 22 '10 at 12:42
  • Good point Chris, just checked NICs and driver versions are 4.8.4.1 according to registry. Have downloaded latest from http://www.broadcom.com/support/ethernet_nic/netxtremeii.php but doesnt appear to have inf file included. Are there differing drivers for Hyper-V at all? Drives are 500GB SATA in RAID1. Thanks for helping out, this is the final irritant I need to resolve...for now anyway! – gary Jul 22 '10 at 13:04
  • Hi Chris, Forget that I managed to update NIC drivers using msi, still no improvement unfortunately, seems very slow which is odd for the best server I have! – gary Jul 22 '10 at 13:56
  • One change that is made to windows with the hyper-v role added or hyper-v core is that its not optimized for this type of operation, a normal server is. Its optimized for guest operations. thats why it's microsofts best practice to have no other roles assigned to a hyper-v server. – tony roth Jul 22 '10 at 14:17
  • Hi Tony, this is core version no other roles are assigned to my knowledge. How would I optimise the server for Hyper-V then? – gary Jul 23 '10 at 05:50

2 Answers2

2

Are you copying files from one VM to another VM on the same Hyper-V host?

If you're doing a physical to physical transfer, one disk/LUN is reading at full speed, and one is writing at full speed. If you are doing VM to VM transfers on the same storage, you can expect half the throughput because the same set of disks are doing both the reads and the writes.

If the VMs are using independent storage, are you using dynamic disks or pass-through? Dynamic disks have a lot lower overhead in Hyper-V R2, but it's not zero. I wouldn't expect a 50% drop in throughput though.

Aaron Friel
  • 588
  • 3
  • 11
0

I assume your only copying host to host, not into a VM like Aaron correctly put the concerns around VM's.

I have tested these scenarios and just by installing Hyper-V role you won't slow a server down. To prove this theory you could remove the role, or reload the box and test throughput before adding role. I suspect something not Hyper-V related is to blame. If you have to regedit after a fresh install then you're likely going down the wrong path.

If it's not drivers then are they the same OS version? SMB is the protocol of file copying from/to Windows and was updated big time in 2008 and again in 2008 R2. You need to ensure your testing is valid based on what OS's are involved. If Win2000/2003 is involved in a file copy transfers will be much slower.

Are you taking disk spindles into account? Newer SAS drivers are faster at I/O then older SCSI or SATA. The number of disks and RAID config matter in file copy as well. Fragmentation can also play a part if using servers that have been around awhile.

You may have formatted disks on the new server with mis-aligned disks. Google can help you find out how to check if your disk formatting is aligned, which if it isn't, can affect performance like you are seeing.

Bret Fisher
  • 3,963
  • 2
  • 20
  • 25