Re: Running PV kernel on XenServer kills all VMs

[Thor Lancelot Simon <tls%panix.com@localhost>]

On Tue, Jun 15, 2010 at 07:38:41PM +0200, Manuel Bouyer wrote:
> 
> I'd say this is a regression. flipping mode should be a bit faster
> for TCP (it was in my benchmarks at last, on a UP system).

I guess it causes more memory bandwidth consumption and possibly more
but reduces interprocessor interrupts for TLB update and possibly
cache synchronization traffic.  It may be that with the huge increase
in memory bandwidth from local memory controllers on all the new CPUs,
copy beats flip.  Certainly it seems like flipping a whole page for a
plan Ethernet frame that can't exceed 1.5K is somewhat wasteful, and
isn't the "has to fit in a page" restriction the reason why our xennet
code cannot do jumbo frames?  That would improve TCP performance between
domains, too...

Is cl still at XenSource?  Maybe he can help us see what the best thing
is to do.

Thor