Re: [Xen-devel] [PATCH 3 of 7] docs: add a document describing the xl cfg file syntax\

[Ian Campbell <Ian.Campbell@xxxxxxxxxx>]

On Thu, 2011-11-10 at 09:54 +0000, Tim Deegan wrote:
> At 08:10 +0000 on 10 Nov (1320912604), Ian Campbell wrote:
> > Tim, George:
> > 
> > Is this broadly accurate? In particular the bit about why one would use
> > the shadow_memory option and the suggestion that it also controls the
> > space used by the HAP overhead.
> > 
> > Cheers,
> > Ian.
> > 
> > On Mon, 2011-11-07 at 15:13 +0000, Ian Campbell wrote:
> > > 
> > > +### Paging
> > > +
> > > +The following options control the mechanisms used to virtualise guest
> > > +memory.  The defaults are selected to give the best results for the
> > > +common case and so you should normally leave these options
> > > +unspecified.
> > > +
> > > + * `hap=BOOLEAN`: Turns "hardware assisted paging" (the use of the
> > > +   hardware' nested page table feature) on or off.  Affects HVM
> > > guests
> > > +   only.  If turned off, Xen will run the guest in "shadow page
> > > table"
> > > +   mode where the guest's page table updates and/or TLB flushes
> > > +   etc. will be emulated.  Use of HAP is the default when available.
> 
> Yep.  Might be worth mentioning that HAP is called EPT and NPT (or RVI)
> by the hardware vendors.

Good idea. Since I had to look them up:
EPT -- Extended Page Tables (Intel)
NPT -- Nested Page Tables (AMD)
RVI -- Rapid Virtualisation Indexing (vmware's name, I think?)

> 
> > > + * `oos=BOOLEAN`: Turns "out of sync pagetables" on or off.  When
> > > +   running in shadow page table mode, the guest's page table updates
> > > +   may be deferred as specified in the Intel/AMD architecture
> > > manuals.
> > > +   However this may expose unexpected bugs in the guest, or find bugs
> > > +   in Xen, so it is possible to disable this feature.  Use of out of
> > > +   sync page tables, when Xen thinks it appropriate, is the default.
> 
> Yep. 
> 
> > > + * `shadow_memory=MBYTES`: Number of megabytes to set aside for
> > > +   shadowing guest pagetable pages (effectively acting as a cache of
> > > +   translated pages) or to use for HAP state. By default this is 1MB
> > > +   per guest vcpu plus 8KB per MB of guest RAM. You should not
> > > +   normally need to adjust this value. However if you are not using
> > > +   hardware assisted paging (i.e. you are using shadow mode) and your
> > > +   guest workload consists of a large number of processes which do
> > > not
> > > +   share address space then increasing this value may improve
> > > +   performance. 
> 
> Actually in the bad case the processes _do_ share address space
> (otherwise the guest would be thrashing for RAM before it rnus out of
> shadow memory).  Maybe 'a very large number of similar processes'?

Sounds fine, thanks!

> 
> Cheers,
> 
> Tim.



_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel