Is there a substantial performance overhead for using Virtual Switch?

Question

In hypervisors such as VMWare products and Hyper-V, there is the option to share the host's network connection with the guests. Can the performance overhead for using this feature be quantified? In terms of:

• CPU usage
• Additional latency
• Bandwidth reduction

How does the overhead of a virtual switch compare to:

1. Using a physical switch with multiple physical adapters

2. Using VLAN's to create multiple virtual adapters (considering the tagging and untagging)

Here is a screenshot of the Hyper-V setting for the feature I'm referring to:

Answer 1

There is a great tech paper on this Here

In the end using virtual switches will add some overhead to the server. This is shown in that tech paper where they run some basic comparisons using different methods. One using virtual switches and one using DirectPath I/O.

In the end to determine the exact overhead that it will take you your case you will have to perform the same tests on your exact configuration, since VM numbers and configuration can greatly affect this.

While there is not quantifiable number. This will give you a good base line even looking at just their comparisons.

On a side note most of the changes in latency are measured in microseconds. I have a private cloud that runs on VMWare and we have no issues with our virtual switches.

Answer 2

Generally, virtualizing all these functions are going to consolidate all the roles to the host resources. Latency differences are going go depend on what your baseline is and resources and performance of your host machine during testing and, of course, production environments.

Also, note that Hyper-V (and ESXi) can offload much of the processing to dedicated I/O cards. So dedicated SR-IOV cards can go addressing any overhead concerns. Or at least add more variables to consider!