perfSONAR User Q&A and Other Discussion

[Aaron Brown <>]

Hi Joseph,

On Feb 20, 2012, at 1:45 PM, Joseph Bernard wrote:

I know it is not recommended to run PS-PS as a virtual machine, but  I am trying to do different things with different network cards on the same box in the same subnet, and Linux doesn't seem to like that.  Has anyone had good luck with network testing using a particular VM platform?

If you want a bandwidth tester, or ping latency tester, Xen or KVM should work just fine for tests up to 1G in bandwidth. There have been reports that NDT could get confused about the causes of performance seen in a virtual machine, but the bandwidth numbers it'd give should be accurate. One-way latency tests are highly dependent on the clock and the latency between the NIC and the measurement daemon (in essence, they end up measuring the latency between the source and destination hosts, the latency inside the source and destination hosts, and the accuracy of the clocks on the source and destination). Virtual machines have higher delays between when the packet arrives on the NIC, and when the userspace daemon receives the packet. Depending on how high that latency is, it can skew the latency numbers considerably. On the clock end of things, virtual machines usually do not have as accurate a clock, and since NTP runs in the hosting machine, and not in the guest, the one-way latency daemon doesn't have any ability to check the accuracy of the clock itself. In the end, the type of virtual machine (KVM vs. Xen) shouldn't matter much as long as your workload is something that would work in a VM environment.

Cheers,

Aaron