The next step is to tune the host operating system both for virtualization and concurrent database oriented client workloads. It is really no different than tuning a stand-alone database server except that the host is one step or level of abstraction removed. You simply need to make sure that you minimize any unnecessary overhead and, thus, maximize the overall performance. The only difference is that CPU, memory and I/O assumptions need to be made a lot more generically since the hardware will be shared across hosted clients. So, optimally, all your tuning efforts should apply relatively well across hosted clients, therefore realizing the double-dipping and triple-dipping effects mentioned earlier.

The chief new complexity is that you now may have multiple databases using some shared resources (much like historically hosting multiple databases on a single box), and thus, need to account for that at some point in your tuning. But each optimization step should occur in its own due time. Tune the virtual server host first. Then tune each of the virtual machines clients (see next chapter) as though you are still implementing single database per server deployments (which technically speaking, could occur at some point even in a virtual world). And finally, tune the host operating system for side effects caused by sharing resources, but do not start here or try to do it all in one pass.

Much like any scientific experiment or database benchmarking, you should only change a single variable per test iteration so that you can properly measure and accord your observed results. I call this technique Incremental Tuning and suggest that it is both the mandatory and only reliable way to correctly optimize any computer system, especially a virtualized host running multiple virtual clients.

Optimize by Subsystem

It is best to think of a virtual host server as being composed of four basic subsystems, which should be the focus of any OS tuning efforts:

• CPU

• Memory

• I/O

• Network

Furthermore, all the above areas should be tuned with the servers purpose in mind hosting multiple virtual clients. The basic idea being that the overall performance can be no better or worse than the sum of its parts. Additionally, all of the above are possibly dynamic in the virtual world, so again making more generic assumptions should generally lead to better aggregate results. This is really nothing more than Incremental Tuning applied at the first granular level of interest: the subsystems.

We will now look at doing just this for both Linux and Windows (note that all these techniques would apply in some fashion to other operating systems such as Sun Solaris, Hewlett Packard's HP-UX or IBMs AIX). In all cases, it is assumed that simply installing a clean basic host operating system is something the user finds both comfortable and familiar. A summarization is included at the end of the chapter as well for future quick reference.

 
This is an excerpt from Oracle on VMWare: Expert tips for database virtualization by Rampant TechPress.