On Thu, Nov 27, 2008 at 10:32 PM, Juha Vuori <[EMAIL PROTECTED]> wrote:

> But another view to the swap problem - or should we say the problem of 
> sharing real memory:
> Can't we live without swap at all and use the cooperative memory management 
> technique?
> You just define enough real memory for each linux instance to handle every 
> peak (well, maybe almost)
> and they just use what they really need and release everything else to the 
> common memory pool.
> You don't have to save free memory for new linux instances - all the memory 
> which is free above VM
> for linuxes can be defined for them and they share it dynamically in the most 
> effective way.

Yep. In theory. You're no doubt aware that we have the instructions in
hardware. Just those nasty details that bother us.
One of the challenges is your common pool. If the Linux guest needs to
obtain a new page from that pool, it would involve the global memory
manager and cause overhead. It would cause a performance penalty for
Linux to give a page back to the pool and ask for one later, as
opposed to just sit on your pages and manage them yourself. This is
not helpful to encourage social behavior. So your ideal hardware would
need a subpool per guest that can be replenished and harvested by the
global memory manager (some of this involves crystal balls predicting
memory demand).
There's also interesting scenarios when only part of your Linux guests
participates in this scheme (because real installations have business
requirements that prevent a big-bang implementation of new kernels
that can do the trick).

> I have not implemented this yet, but looking forward to do that.
> Am i too enthusiastic?
> CMM _must_ have a problematics of its own ... :)

My prime concern is the lack of instrumentation and controls (which is
not easy to fix with things in hardware now). Contrasting popular
belief, it's probably harder to drive it without meters and controls.
And while the Linux side of this can be understood from the source (if
you try hard enough) the hardware and hypervisor side of it has not
been disclosed. The best we can do is guess at what is happening
there. You can spend entire evenings in a hotel bar discussing this
based on assumptions and belief, as long as you're not corrected by
seeing actual data. You can even come up with synthetic workload that
would exploit it very well. The presentations I've seen about CMM do
one or both ;-)

Rob
--
Rob van der Heij
Velocity Software
http://velocitysoftware.com/

----------------------------------------------------------------------
For LINUX-390 subscribe / signoff / archive access instructions,
send email to [EMAIL PROTECTED] with the message: INFO LINUX-390 or visit
http://www.marist.edu/htbin/wlvindex?LINUX-390

Reply via email to