I came across an interesting problem the other day while I was doing some unrelated troubleshooting on one of my Hyper-V guests.
The symptoms were that my Windows Server 2003 machine would return very strange results when pinging hosts, both internally and externally, such as returning all four responses within about half a second, yet measuring them at over 3000ms (which means they should have timed out, rather than given me a reading in milliseconds) as well as occasionally providing negative values for response times.
Obviously the results were completely inaccurate, but I couldn’t work out why the issue was only happening on a handful (not all) Hyper-V guests running Windows Server 2003 and none on Server 2008.
Turns out that this is an issue if all of the following are true:
- You are running an operating system prior to Windows Vista or Windows Server 2008
- You are running the current implementation of Microsoft Hyper-V (i.e. at the time of writing)
- You have presented multiple processors to the Hyper-V guest
The issue occurs because the multiprocessor HAL in Hyper-V causes the guest’s operating system Time Stamp Counter (TSC) to skew. According to this blog the problem wouldn’t ordinarily occur if you were running Windows Server 2003 with SP2 unless the BIOS check fails to determine if the TSC should be used. More specifically, if I understand correctly the issue occurs because the processors (or cores, if we’re talking about a single multicore processor) are not in sync with each other, which produces sporadic out-of-time results where time sensitive operations (such as ping responses) are in use.
The resolution is to force the guest to use the PM timer instead of the TSC, by adding /USEPMTIMER in the boot.ini file and then restart. You can easily test this by running a ping -t to a host and checking for drastically abnormal results.