While talking with people about virtualization, I have noticed that many think they can just install ESX on any available server and use all of that server’s resources. This article discusses two of the most common misconceptions I have encountered:

“We will install ESX on an HP superdome!”

VMware products (including ESX 2.x and 3.0) run on x86 architecture only, not on Itanium. HP superdomes use Itanium processors. ESX is meant to virtualize x86 operating systems such as NT 4, Windows 2000/2003, Linux, etc… Doing this on Itanium would probably mean big performance hits.

Note that is is possible to run x64 operating systems such as Windows Server 2003 x64 on ESX 3.0 (experimental support). But in order to do that, you need to install ESX 3 on an x64 capable host. On Intel, VT support is needed as well.

“We will install ESX on a system with 32 CPUs and use them all!”

Some customers want a scale up solution with big boxes running many virtual machines. There are some x86–based systems such as the IBM x460 that support 32 Xeon CPUs. The x460 is actually a four-way capable server but it is expandable by adding additional nodes. The additional nodes can be other x460 systems or the MXE-460 modular expansion enclosure. It is possible to build a physical server with 32 CPUs, 512GB of RAM and 48 PCI-X slots. Take into account that these 32 CPUs can be dual-core with hyper-threading support which means 128 logical CPUs.

Clearly, such a system is beyond the capabilities of ESX. ESX 2.x supports 16 logical CPUs. This means that with the x460, you can build a 16–way system with single-core CPUs and hyper-threading turned off, or an 8–way system with dual-core CPUs and hyper-threading turned off. ESX 2.x also supports a maximum of 80 virtual CPUs. This means you can add a maximum of 80 single CPU virtual machines, or 40 dual CPU, or a mix.

ESX 3.0 increases the number of supported logical CPUs to 32. Taking the x460 as an example, you will be able to go up to 8 nodes with 32 single-core CPUs with hyper-threading turned off. Or you could use 8 dual-core CPUs with hyper-threading turned on (2 nodes). Of course, the 32 single-core CPUs would have better performance.