> > Hmmm... if the goal is just to run Linux I would think UML or > alternatives > would be more efficent. VMware is very nice, but if they're > all Linux... > might as well gain back some performance (VMware is resource > expensive).
Keep in mind also that there are 3 generations of VMWare -- the desktop one, which is not suitable for production use, GSX (which retains the "hosted" concept using an native OS and simulation), and ESX (which runs on the bare metal and provides a lot more function). GSX and ESX are the only versions which make sense to be used for server hosting. ESX is the only version that offers >1 CPU and the ability to cap resource utilization of a virtual machine. > Not sure if 6 per "engine" is reasonable with VMware, you'll > need a LOT > of memory (which presents its own set of problems with both the Intel > platform and VMware). It's not. Recommended load per engine is no more than 4 virtual machines per physical CPU, and empirical testing indicates that 3 is a better number (to allow for bursting). To get 24 virtual machines with predictable production scale performance, you're looking at 8 way Intel systems running ESX. That said, GSX and ESX have come a long way -- you can now dedicate resources to specific virtual machines, and PCMCIA support is much better in the advanced versions. As mentioned above, ESX now provides >1 CPU emulation, and the scripting interface has improved greatly in the last version (two or three more revs and it'll be almost as good as the LPAR definitions screens on a 3090-100E circa 1988...8-)). > Now.. if VMware supported large scale Opteron platforms!! > (just dreaming at > the moment) ... might even challenge z/VM (?) (well.. not management > wise... but scalability wise). Not really. IA32 CPUs are hard to emulate -- lots of stupid design tricks. IA64 and the newer Opteron boxes are a lot less stupid. Wrt to large scale platforms, I've seen a 16 way x440 running VMWare-ESX. It makes a nice hosted environment, but it's also up in the $70-80K range when all is said and done (if you go with that level of consolidation, you need the full no-excuses redundancy features). -- db >
