Re: [Qemu-devel] [Qemu-ppc] [RFC PATCH] smp: autodetect numbers of threads per core

[Alexander Graf]

On 15.11.13 17:58, Alexey Kardashevskiy wrote:
> On 16.11.2013 0:15, Alexander Graf wrote:
> > Am 15.11.2013 um 00:12 schrieb Alexey Kardashevskiy <address@hidden>:
> >
> > > At the moment only a whole CPU core can be assigned to a KVM. Since
> > > POWER7/8 support several threads per core, we want all threads of a core
> > > to go to the same KVM so every time we run QEMU with -enable-kvm on
> > > POWER, we have to add -smp X,threads=(4|8)" (4 for POWER7 and
> > > 8 for POWER8).
> > >
> > > This patch tries to read smp_threads number from an accelerator and
> > > falls back to the default value (1) if the accelerator did not care
> > > to change it.
> > >
> > > Signed-off-by: Alexey Kardashevskiy <address@hidden>
> > > ---
> > >
> > > (!!!)
> > >
> > > The usual question - what would be the normal way of doing this?
> > > What does this patch break? I cannot think of anything right now.
> > Is this really what the user wants? On p7 you can run in no-smt, smt2
> > and smt4 mode. Today we simply default to no-smt. Changing defaults
> > is usually a bad thing.
>
> Defaulting to 1 thread on P7 is a bad thing (other threads stay unused -
> what is good about this?) and the only reason which I know why it is
> still threads=1 is that it is hard to make a patch for upstream to
> change this default.

threads=1 improves single-thread performance significantly. The thread 
itself is faster when it runs in SMT1 mode. Also we don't have to kick 
other threads out of the guest context, making every guest/host 
transition faster.

Overall, it's really just a random default. I'm not sure it makes a lot 
of sense to change it.

However, could we be really smart here? How does ppc64_cpu --smt=off 
work? It only turns off the unused vcpus, right? Is there any way we 
could actually not even enter unused vcpus at all? Then we could indeed 
always expose the maximum available number of threads to the guest and 
let that one decide what to do.


Alex