Port-xen archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

Re: 2nd round of xen benchmarking



On Mon, Aug 19, 2024 at 05:07:46PM +0200, Manuel Bouyer wrote:
> On Mon, Aug 05, 2024 at 06:35:09PM -0400, Greg Troxel wrote:
> > I have swapped out one ancient system for another.  This one has VMX
> > feature, but doesn't have the more modern virtualization features.
> > 
> > The basic benchmark is building devel/m4, no ccache, MAKE_JOBS=1, work
> > on tmpfs.  1 vcpu for dom0 and domU.  (Multiple CPUs on bare metal, but
> > MAKE_JOBS=1.)
> > 
> > Overall, it looks like (talking about times) commpared to bare metal times:
> > 
> >   dom0 is 2x
> >   pv (or pvh/pvshim) domU is about 3x
> > 
> >   n9-amd64: hvm is 19x (!!!!)
> > 
> >   n10-amd64: pvh is 4.5 slower than pv, and pvhvm is 4x slower than pv
> > 
> > It remains a really good question if the newer CPUs will lead to reduced
> > slowdowns.   As it is, pvh is a win for n10-i386, and other than that,
> > pv is best, sometimes by a huge amount.
> 
> For PVH to be a win, my guess is that you really want a CPU with EPT. 
> Otherwise each page table operation (and pkgsrc does a lot of fork/exec)
> causes a trap to the hypervisor. PV page table operations is cheaper.

I've run a similar benchmarck:
/usr/bin/time sh -c "make; make clean" in pkgsrc/devel/m4, with
WRKOBJDIR in tmpfs make MAKE_JOBS=1
This is on a Xeon E5-2637, which does have EPT, the domU has 8 vcpus and
8GB RAM.

PV (netbsd-XEN3_DOMU.gz from the NetBSD 10.0 release)
      155.66 real        64.30 user        91.96 sys
      153.36 real        63.57 user        92.71 sys
      153.79 real        64.68 user        92.20 sys
      153.77 real        64.89 user        91.99 sys

PVH (netbsd-GENERIC.gz from the NetBSD 10.0 release)
       64.08 real        38.36 user        19.98 sys
       61.28 real        38.80 user        20.05 sys
       61.31 real        38.79 user        20.07 sys
       61.38 real        39.08 user        19.83 sys

As expected, with EPT PVH gives a good performance boost over PV.
I can't test bare metal as this is a production system running several VMs.

-- 
Manuel Bouyer <bouyer%antioche.eu.org@localhost>
     NetBSD: 26 ans d'experience feront toujours la difference
--


Home | Main Index | Thread Index | Old Index