Virtual Machine Performance Demystified

1. Everywhere he talks about "memory", he really means "privileged virtual memory operations." Flushing the TLB, altering the virtual to physical mappings, changing address spaces, etc. Ordinary loads and stores have always been directly executed at native speed. He also merges "apache" (lots of context switches) and "compile" (lots of process creation and page faults) into the odd-sounding "apache compile").

2. In my opinion, paravirtualization was first a hack, and only retrospectively framed as a performance optimization. By hacking up Linux, the Xen folks were able to get a viable system working much more easily than would have been possible for unmodified kernels. The notional perf advantages of this approach have, IMHO, not been convincingly shown, and RVI/EPT hardware impressively raise the bar for ever demonstrating it.

3. Although I worked on it, I would be very shy about claiming "hypervisor" as a VMware invention. First, it is not an entirely accurate description of ESX, which keeps a much more rigid separation of the "monitor" (responsible for providing virtual CPUs, memory management hardware, and devices) from the "vmkernel" (responsible for driving IO and keeping VMs out of one another's way) than most people imagine when they hear the term "hypervisor". But most importantly, I first remember hearing the term applied to kernels from the first virtualization golden age, in the 1970's. We have stood on these giants' shoulders enough without claiming their inventions for ourselves.  

FWIW IO performance on Joyent is abysmal, and they supposedly use Sun's latest virtualization technology.

It sounds like the problem is not with the technology but with how it's managed... It's easy to partition memory and CPU resources, but not as easy to partition IO resources.

I'm still waiting for virtual hosts that are paired with their own dedicated solid state storage rather than sharing one hardware IO system or using slow network storage.