Re: [patch 0/3] kvm tool: Serial emulation overhaul

[Thomas Gleixner <tglx@linutronix.de>]

On Sun, 11 Dec 2011, Ingo Molnar wrote:
> * Thomas Gleixner <tglx@linutronix.de> wrote:
> 
> > On Sun, 11 Dec 2011, Ingo Molnar wrote:
> 
> > > In theory a kvm driven serial console should be extremely 
> > > fast, much faster than a real serial console, basically as 
> > > fast as a local console. Yet this still does not seem to be 
> > > the case.
> > > 
> > > In fact even ssh-ing in to a box over Wifi and running the 
> > > above top session is undistinguishable from top running in a 
> > > local console. So IMO it cannot be virtualization overhead - 
> > > there must still be some delay or serious lack of buffering 
> > > somewhere.
> > 
> > Well, the difference between ssh and serial is, that ssh can 
> > pack 1.5k worth of data into one frame, while serial has to 
> > send it piecewise. And the emulation has to trap into kvm tool 
> > for each tx byte, which doesnt help either. We cannot do much 
> > with buffering on the kvm tool side as we have no clue how 
> > much consecutive data will come in. That's why there is a 
> > virtual console, which has the disadvantage that you cant see 
> > the early boot messages.
> 
> Okay, but look at it from another angle: the top output i 
> generate is about 300k characters. 5000 msecs to execute it 
> means 16 usecs overhead per character - or about 50k cycles - on 
> a top of the class x86 CPU.
> 
> 50k cycles for every single byte. And as a user i notice that 
> first hand.

Well, you can't do anything about it simply because it's doing full
hardware emulation which goes all the way to user space and back for
each inb/outb.

virtio_console was written to avoid that overhead.

Thanks,

	tglx