This patch is intended to start a slightly larger discussion about
our plans for the CentOS CI environment going forward.

At the moment, we have active builders for

  CentOS 7
  Debian 9
  Debian 10
  Fedora 30
  Fedora 31
  Fedora Rawhide
  FreeBSD 11
  FreeBSD 12

but we don't have builder for

  Debian sid
  FreeBSD -CURRENT
  Ubuntu 16.04
  Ubuntu 18.04

despite them being fully supported in the libvirt-jenkins-ci
repository.

This makes sense for sid and -CURRENT, since the former covers the
same "freshest Linux packages" angle that Rawhide already takes care
of and the latter is often broken and not trivial to keep updated;
both Ubuntu targets, however, should IMHO be part of the CentOS CI
environment. Hence this series :)

Moreover, we're in the process of adding

  CentOS 8
  openSUSE Leap 15.1
  openSUSE Tumbleweed

as targets, of which the first two should also IMHO be added as they
would provide useful additional coverage.

The only reason why I'm even questioning whether this should be done
is capacity for the hypervisor host: the machine we're running all
builders on has

  CPUs: 8
  Memory: 32 GiB
  Storage: 450 GiB

and each of the guests is configured to use

  CPUs: 2
  Memory: 2 GiB
  Storage: 20 GiB

So while we're good, and actually have plenty of room to grow, on
the memory and storage front, we're already overcommitting our CPUs
pretty significantly, which I guess is at least part of the reason
why builds take so long.

Can we afford to add 50% more load on the machine without making it
unusable? I don't know. But I think it would be worthwhile to at
least try and see how it handles an additional 25%, which is exactly
what this series does.

In my opinion, as long as the machine can keep up with demand and not
end up in a situation where it starts accumulating backlog jobs, it's
fine if builds take longer: developers who want to run a relatively
quick smoke test before posting patches can use Travis CI or 'make
ci-check@...' locally for the purpose, whereas the role of CentOS CI
in my eyes is to try and catch as many issues as possible after merge
so that they don't end up in a release. But I realize others might
see it differently, hence this lenghty cover letter :)

Andrea Bolognani (1):
  jenkins: Start building on Ubuntu

 jenkins/jobs/defaults.yaml            | 2 ++
 jenkins/projects/libvirt-dbus.yaml    | 1 +
 jenkins/projects/libvirt-sandbox.yaml | 2 ++
 jenkins/projects/libvirt-tck.yaml     | 2 ++
 jenkins/projects/libvirt.yaml         | 2 ++
 jenkins/projects/virt-manager.yaml    | 2 ++
 6 files changed, 11 insertions(+)

-- 
2.23.0

--
libvir-list mailing list
libvir-list@redhat.com
https://www.redhat.com/mailman/listinfo/libvir-list

On Tue, Dec 10, 2019 at 02:54:22PM +0100, Andrea Bolognani wrote:
> This patch is intended to start a slightly larger discussion about
> our plans for the CentOS CI environment going forward.
> 
> At the moment, we have active builders for
> 
>   CentOS 7
>   Debian 9
>   Debian 10
>   Fedora 30
>   Fedora 31
>   Fedora Rawhide
>   FreeBSD 11
>   FreeBSD 12
> 
> but we don't have builder for
> 
>   Debian sid
>   FreeBSD -CURRENT
>   Ubuntu 16.04
>   Ubuntu 18.04
> 
> despite them being fully supported in the libvirt-jenkins-ci
> repository.
> 
> This makes sense for sid and -CURRENT, since the former covers the
> same "freshest Linux packages" angle that Rawhide already takes care
> of and the latter is often broken and not trivial to keep updated;
> both Ubuntu targets, however, should IMHO be part of the CentOS CI
> environment. Hence this series :)
> 
> Moreover, we're in the process of adding
> 
>   CentOS 8
>   openSUSE Leap 15.1
>   openSUSE Tumbleweed
> 
> as targets, of which the first two should also IMHO be added as they
> would provide useful additional coverage.
> 
> The only reason why I'm even questioning whether this should be done
> is capacity for the hypervisor host: the machine we're running all
> builders on has
> 
>   CPUs: 8
>   Memory: 32 GiB
>   Storage: 450 GiB
> 
> and each of the guests is configured to use
> 
>   CPUs: 2
>   Memory: 2 GiB
>   Storage: 20 GiB
> 
> So while we're good, and actually have plenty of room to grow, on
> the memory and storage front, we're already overcommitting our CPUs
> pretty significantly, which I guess is at least part of the reason
> why builds take so long.

NB the memory that's free is not really free - it is being usefull
as I/O cache for the VM disks. So more VMs will reduce I/O cache.
Whether that will actually impact us I don't know though.

More importantly though, AFAICT, those are not 8 real CPUs.

virsh nodeinfo reports 8 cores, but virsh capabilities
reports it as a 1 socket, 4 core, 2 thread CPU.

IOW we haven't really got 8 CPUs, more like equivalent of 5 CPUs.
as HT only really gives a x1.3 boost in best case, and I suspect
builds are not likely to be hitting the best case.


> Can we afford to add 50% more load on the machine without making it
> unusable? I don't know. But I think it would be worthwhile to at
> least try and see how it handles an additional 25%, which is exactly
> what this series does.

Giving it a try is ok I guess.

I expect there's probably more we can do to optimize the setup
too.

For example, what actual features of qcow2 are we using ? We're
not snapshotting VMs, we don't need grow-on-demand allocation.
AFACT we're paying the performance cost of qcow2 (l1/l2 table
lookups & metadata caching), for no reason. Switch the VMs to
fully pre-allocated raw files may improve I/O performance.
Raw LVM VGs would be even better but that will be painful
to setup given the host install setup.

I also wonder if we have the optimal aio setting for disks,
as there's nothing in the XML.

We could consider using cache=unsafe for VMs, though for
that I think we'd want to separate off a separate disk
for /home/jenkins so that if there was a host OS crash,
we wouldn't have to rebuild the entire VMs - just throw
away the data disk & recreate.

Since we've got plenty of RAM, another obvious thing would be
to turn on huge pages and use them for all guest RAM. This may
well have a very significant performance boost from reducing
CPU overhead which is our biggest bottleneck.

Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

--
libvir-list mailing list
libvir-list@redhat.com
https://www.redhat.com/mailman/listinfo/libvir-list

On Tue, 2019-12-10 at 14:54 +0000, Daniel P. Berrangé wrote:
> On Tue, Dec 10, 2019 at 02:54:22PM +0100, Andrea Bolognani wrote:
> > The only reason why I'm even questioning whether this should be done
> > is capacity for the hypervisor host: the machine we're running all
> > builders on has
> > 
> >   CPUs: 8
> >   Memory: 32 GiB
> >   Storage: 450 GiB
> > 
> > and each of the guests is configured to use
> > 
> >   CPUs: 2
> >   Memory: 2 GiB
> >   Storage: 20 GiB
> > 
> > So while we're good, and actually have plenty of room to grow, on
> > the memory and storage front, we're already overcommitting our CPUs
> > pretty significantly, which I guess is at least part of the reason
> > why builds take so long.
> 
> NB the memory that's free is not really free - it is being usefull
> as I/O cache for the VM disks. So more VMs will reduce I/O cache.
> Whether that will actually impact us I don't know though.

Oh yeah, I'm aware of that. But like you, I'm unsure about the exact
impact that makes.

We could arguably reduce the amount of memory assigned to guests to
1 GiB: building libvirt and friends is not exactly memory-intensive,
and we could potentially benefit from the additionaly I/O cache or
lessen the blow caused by adding more VMs.

> More importantly though, AFAICT, those are not 8 real CPUs.
> 
> virsh nodeinfo reports 8 cores, but virsh capabilities
> reports it as a 1 socket, 4 core, 2 thread CPU.
> 
> IOW we haven't really got 8 CPUs, more like equivalent of 5 CPUs.
> as HT only really gives a x1.3 boost in best case, and I suspect
> builds are not likely to be hitting the best case.

That's also true.

> > Can we afford to add 50% more load on the machine without making it
> > unusable? I don't know. But I think it would be worthwhile to at
> > least try and see how it handles an additional 25%, which is exactly
> > what this series does.
> 
> Giving it a try is ok I guess.

Can I have an R-b then? O:-)

> I expect there's probably more we can do to optimize the setup
> too.
> 
> For example, what actual features of qcow2 are we using ? We're
> not snapshotting VMs, we don't need grow-on-demand allocation.
> AFACT we're paying the performance cost of qcow2 (l1/l2 table
> lookups & metadata caching), for no reason. Switch the VMs to
> fully pre-allocated raw files may improve I/O performance.
> Raw LVM VGs would be even better but that will be painful
> to setup given the host install setup.
> 
> I also wonder if we have the optimal aio setting for disks,
> as there's nothing in the XML.
> 
> We could consider using cache=unsafe for VMs, though for
> that I think we'd want to separate off a separate disk
> for /home/jenkins so that if there was a host OS crash,
> we wouldn't have to rebuild the entire VMs - just throw
> away the data disk & recreate.

The home directory for the jenkins user contains some configuration
as well, so you'd have to run 'lcitool update' anyway after attaching
the new disk... At that point, it might be less work overall to just
rebuild the entire VM from scratch. We have only seen a couple of
unexpected host shutdowns so far, and all were caused by hardware
issues that resulted in a multi-day downtime anyway, so the overhead
of reinstalling the various guest OS' would not be the dealbreaker.

> Since we've got plenty of RAM, another obvious thing would be
> to turn on huge pages and use them for all guest RAM. This may
> well have a very significant performance boost from reducing
> CPU overhead which is our biggest bottleneck.

Yeah, all of the above sounds like it could help, but I'm not really
well versed on the performance tuning front so I wouldn't know where
to start and even how to properly measure the impact of each change.

We also have to keep in mind that, while the CentOS CI environment
is the main consumer of the repository, we want it to be possible for
any developer to deploy builders locally relatively painlessly, so
things like hugepages and LVM usage would definitely have to be
optional if we introduced them.

-- 
Andrea Bolognani / Red Hat / Virtualization

--
libvir-list mailing list
libvir-list@redhat.com
https://www.redhat.com/mailman/listinfo/libvir-list