Re: [PATCH 0/8] KVM updates for 2.6.20-rc2

[Avi Kivity <avi-atKUWr5tajBWk0Htik3J/w@public.gmane.org>]

Ingo Molnar wrote:
> * Avi Kivity <avi-atKUWr5tajBWk0Htik3J/w@public.gmane.org> wrote:
>
>   
>> Talking about the scheduler (and to the scheduler's author :), it 
>> would be nice to hook the migration weight algorithm too.  kvm guests 
>> are considerably more expensive to migrate, at least on intel.
>>     
>
> what do you mean precisely, and where does the cost come from? To me it 
> seems the cost of loading/saving a VMX context on Intel is the biggest 
> cost, and that's roughly the same cost for VM exits/enters, regardless 
> of what the scheduler does. What am i missing?
>   

There are three types of vmx context switches.

1. The most common switch is a vm entry/exit pair (vmresume 
instruction).  The context is already loaded into the cpu (via a 
vmptrld, below), and we enter into a preloaded VM.  In addition, 
register access implicitly addresses the loaded context.

2. The second most common switch is loading another context.  If we have 
several VMs on a uniprocessor machine, the execution sequence looks like 
this:

  vmptrld vm1

  ... access context registers (vmread/vmwrite)
  vmresume
  ... access context registers (vmread/vmwrite)

  ... access context registers (vmread/vmwrite)
  vmresume
  ... access context registers (vmread/vmwrite)

  ... access context registers (vmread/vmwrite)
  vmresume
  ... access context registers (vmread/vmwrite)

  Linux context switch to another kvm process

  vmptrld vm2

  ... access context registers (vmread/vmwrite)
  vmresume
  ... access context registers (vmread/vmwrite)

  ... access context registers (vmread/vmwrite)
  vmresume
  ... access context registers (vmread/vmwrite)

  Linux context switch to back to original process

  vmptrld vm1

Linux context switches to non-kvm processes and back need no special 
handling.

The processor caches the current VM context, and possibly more than one.

3. The most expensive vmx context switch involves cpu migration:

cpu 0:  vmclear vm1 /* decache vmx context into memory */
cpu 1:  vmptrld vm1
            ... vm register access
            vmlaunch

The vmlaunch instruction, like vmresume, causes a VM entry, but is 
documented to be significantly more expensive.  It is required after a 
vmclear.

Currently, the vmclear is performed by an ipi, because we can only 
detect migration after the fact.  However, if we enlist the migration 
threads, we can vmclear before the process has left the cpu.


-- 
error compiling committee.c: too many arguments to function


-------------------------------------------------------------------------
Take Surveys. Earn Cash. Influence the Future of IT
Join SourceForge.net's Techsay panel and you'll get the chance to share your
opinions on IT & business topics through brief surveys - and earn cash
http://www.techsay.com/default.php?page=join.php&p=sourceforge&CID=DEVDEV