| |
|
 |
|
 |
|
|
|
| |
|
|
xen-users
RE: [Xen-users] RE: [Xen-devel] Modular Xen
> -----Original Message-----
> From: Goswin von Brederlow
> [mailto:brederlo@xxxxxxxxxxxxxxxxxxxxxxxxxxx]
> Sent: 19 January 2007 11:27
> To: Petersson, Mats
> Cc: xen-users@xxxxxxxxxxxxxxxxxxx
> Subject: Re: [Xen-users] RE: [Xen-devel] Modular Xen
>
> "Petersson, Mats" <Mats.Petersson@xxxxxxx> writes:
>
> > Resent with xen-users list as recipient, rather than
> directly to Daniel.
> >
> >> -----Original Message-----
> >> From: xen-devel-bounces@xxxxxxxxxxxxxxxxxxx
> >> [mailto:xen-devel-bounces@xxxxxxxxxxxxxxxxxxx] On Behalf Of
> >> Daniel Stodden
> >> Sent: 18 January 2007 15:38
> >> To: Xen Developers
> >> Subject: Re: [Xen-devel] Modular Xen
> >>
> >> On Mon, 2007-01-15 at 19:30 +0200, David Pilger wrote:
> >> > It seems to me that Xen is full of stuff that aren't
> beneficial for
> >> > everybody, this includes NUMA, ACM and other stuff that certain
> >> > companies try to introduce...
> >>
> >> can anyone explain to me what the current numa-bashing is
> all about?
> >> i'm running a lot of my stuff on 4-way opteron servers, these
> >> system are
> >> far from uncommon these days and certainly beyond a mere
> 'introduction
> >> attempt'.
> >>
> >> are there any numbers regarding the code size/performance
> penalty on
> >> systems limited to a single memory node? i'd really be interested.
> >
> > I think this is the wrong thread for this subject (the one
> titled SMP & NUMA something or other started by Emmanuel
> Ackaroy (excuse spelling) is the right one...).
> >
> > But to sumarize the discussion, I don't think anyone is
> denying the need to have Xen understand NUMA. However, there
> are plenty of reasons that NUMA won't necessarily give any
> noticable benefit for the typical Xen system - for example,
> the default configuration is that all guests run on ANY
> available CPU, which means that allocating it's memory on any
> particular CPU will be just as right (or wrong) as allocating
> it in a random or scattered fashion.
> >
> > I completely understand where Emmanuel is coming from when
> he is "negative" about the benefits of NUMA in a Xen system.
> There are places with concrete and obvious benefits, but
> there's also a great risk that things just run slower if it's
> implemented in the wrong way (or USED in the wrong way).
> >
> > So there's some need to research the subject before adding
> the features into the Xen kernel. [The same applies to the
> Linux kernel of course].
>
> Just so I know, does XEN pass on the NUMA infos from the allocated
> memory to the domU linux kernel?
Not at present. The current Xen kernel doesn't have NUMA support.
>
> Say I do have NUMA and 2 cpus and start a domU with 1G ram and 2
> vcpus. Does it allocate 512MB per NUMA node and tell linux that?
As above. And more importantly, even if it did have NUMA, it will have
to allocate memory as per the PHYSICAL CPU architecture allows, which
may not allow 512MB per NODE in a 2-node system.
>
> And do vcpus stay mapped to the same real cpu in case cpus == cpus or
> do they randomly switch (making numa optimizations impossible)?
If you haven't restricted the CPU's in the virtual machine, then the
vcpu can and will switch to any available physical cpu, which of course
means that NUMA becomes a bit harder to handle.
--
Mats
>
> MfG
> Goswin
>
>
>
_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-users
|
|
|
| |
|
Home