Re: [Plex86-devel] Re: [Plex86-developers] Performance characteristics

[Daniel Phillips]

On Thu 20 Feb 03 07:06, niraj tolia wrote:
> Actually VMWare gets very close to raw performance for operations that do
> not need to be trapped by the VM. In other words, I have observed VMWare to
> be quite fast for certain applications. For example, VMWare is fast enough
> to saturate 100 Mbit connections.

This is true of UML as well, when it is not doing, e.g., block IO or 
rearranging memory.  The bottom line performance is going to depend on how 
often the application hits the slow paths, and how slow the slow paths are.  
My impression  of Kevin's current work is that it strips away the horribly 
slow paths, and uses only the virtualization techniques that the processor is 
capable of handling efficiently.

As Kevin keeps saying, the big difference between UML and the newly 
simplified Plex86 is that Plex86 is not a port.  It will be much less 
sensitive to rapidly changing internals in the guest kernel.  As it is, 
every new 2.5 release tends to break UML and Jeff Dike energetically puts it 
back together again, causing UML's availability to lag Linus's releases by a 
days or two, never mind Linus's BK version.  Plex86 on the other hand should 
just work, giving it a big advantage for bleeding edge kernel hacking.

Api changes in the stable kernel may break Plex86's kernel module from time 
to time, but that's no big deal, since a kernel developer can continue using 
the old host that works.

Regards,

Daniel