slow network perf after coldclone of dc

[joergriether]

hello,

yesterday we did a clod clone of a dc (w2003 sp2) which was equipped with an old icp gtd raid controller, cold clone was no problem after giving p2v cd the disk driver at boot time. p2v was done directly vc and to one of our esx 3.5 machines. no problems at all. the physical machine was a dual cpu system therefor the virtual machine is also a dual cpu and the multicpu hal is logically loaded. now what makes me thinking is the fact that the network performance of the new virtual machine is poor. If i transfer large files from other virtual machines (also w2003 sp2), speed is about 50% of the virtual gig nic, about 45.000kb/s, if i transfer from this new vm, speed is about 5% of virtual gig nic, about 4.500kb/s. Anyone know why that would happen? vmware tools are the latest, old nic was removed from hardware inventory with the wellknown show nonpresent devives=1 stuff, all worked well. But the network speed is poor. Could it be somehow connected to the two cpu system? any other thoughts?

best regards

Joerg

[AWo]

It is only a guess, how much physical CPU's are in your host?

[joergriether]

the esx has two quad core xeon 2.6, so we have 8 cores in the host.

[AWo]

That might lead to a overall bad performance, as it is symmetric multiprocessing. That means the guest has to wait for both CPU's to be in sync before it can go on. As other guests are also running on this CPU's this leads to delays. A rule of thumb says you should have double the number of physical processors than you assign to the guests.

[joergriether]

negative. there are only three vms at this time, two have only single vcpu, the new one (the slow one) has two vcpu´s, makes overall 4 vcpu´s and i have 8 pcpu´s.

other ideas?

[AWo]

Read oreeh's comment on this: http://communities.vmware.com/message/952130#952130

and Dave Mishchenko's comment: http://communities.vmware.com/message/951242#951242

You said you have two physical CPU's, not cores. Than this bottleneck applies. Even if the two other guest's only have one vCPU, they will make it harder for the vSMP guest with two vCPU's to schedule the CPU's .

I don't know if this will cause also the slow network traffic in your case, but if you only have two physical CPU's/socket's, you may want to look into this, also.

[joergriether]

to verify, i just installed another w2003 vm mit 2vcpu´s to the same esx server, network speed is like charm, faster than the single vcpu i tested before, nearly 60% usage, 58.000kbp/s. So i guess the problem lies somewhere else.

[depping]

Well AWo with the "relaxed co-scheduling" esx uses these days it

shouldn't actually hurt you that much. Only when there's an app that is

actually specifically using both processors. 9 out of 10 times this

isn't the case and the second processor is just in an idle loop.

Relaxed co-scheduling recognises this idle loop and doesn't co-schedule

the second proc.

Still I agree, cut down on the vcpu, it's pointless.

I would uninstall the tools and install them again, just to be sure! Also check if there is any vendor related software left on the machine and deinstall this software.

Duncan

My virtualisation blog:

If you find this information useful, please award points for "correct" or "helpful".

[kjb007]

Did you install vm tools during the clone or after? I would uninstall them, reboot, and then install fresh.

-KjB

[AWo]

Try to disable TCP chimney and DisableDOS on host OS.

For background informatuion look here: http://support.microsoft.com/kb/945977

The command. "Netsh int ip set chimney DISABLED"

And here: http://support.microsoft.com/kb/898468/en-us

Edit the "HKEY_LOCAL_MACHINE\System\CurrentControlSet\Services\LanManServer\Parameters" entry by adding the "DisableDos" DWORD Key with a value of "1".