[ovirt-devel] [VDSM] rhel6.5 vs rhel7 performance in sampling

[Nir Soffer]

----- Original Message -----
> From: "Francesco Romani" <fromani at redhat.com>
> To: devel at ovirt.org
> Sent: Thursday, July 31, 2014 2:49:27 PM
> Subject: [ovirt-devel] [VDSM] rhel6.5 vs rhel7 performance in sampling
> 
> Hi everyone,
> 
> since the beginning of this week, saved the test day of course,
> I'm doing tests and benchmarking on the scalable_sampling
> patchset from me and Nir.
> 
> http://gerrit.ovirt.org/#/q/status:open+project:vdsm+branch:master+topic:scalable_sampling,n,z
> (plus http://gerrit.ovirt.org/#/c/30751/)
> 
> I'm collecting quite a lot of data and still digesting them, but
> I think it is safe to present a couple of pictures that shows
> hoe much better rhel7 behaves with respect to rhel6.5.
> 
> platform:
> on minidell (core i7, 4 cores, 8 threads, 16 GB Ram)
> engine is a recent 3.5 ~beta2; storage domain is ISCSI
> VDSM master from Jul 24 (not much as changed meantime)
> hypervisor runs stock RHEL6.5 or stock RHEL7.0
> 
> test scenario:
> boot 100 VMs (each 1cpu, 64 MB ram, 1GB disk)
> start sampling on hypervisor using stats from /proc each 1/2 second
> wait 10 minutes
> stop sampling, save sample points
> stop VMS
> 
> what we want to see here is how much resources VDSM alone wastes
> in sampling during the common operations.
> Of course the less is better.
> 
> Nothing fancy here, just the differences from RHEL6.5 to RHEL7.
> See the attached graphs:
> * baseline_{cpu,mem}.png is RHEL 6.5
> * baseline_{cpu,mem}_7.png is RHEL 7.0
> RHEL7 is so *much* better, switching the hypervisor host is by far
> the biggest (and easiest) gain we achieved so far.

Are these showing the current vdsm code, before the scalable_sampling
branch?

> 
> coming soon:
> - discussion about the impact of the scalable_sampling topic branch

I want to see the numbers the new graphs

> - move to 200, 300 ... 500 VMs on a single host.
> 
> PS: a couple of updates about specific points (Michal, Nir)
> - still no luck with yappi on RHEL7, managed to do the above bench
>   with 10 VMs and collected a profile.

Can you share the profile with us?

> - 30751 is a win, same of the last changes of
> http://gerrit.ovirt.org/#/c/29980/
>   (revs 12, 13)

More details?

> - will tick 'V' on http://gerrit.ovirt.org/#/c/29607/ once the last checks
>   are passed (ETA late today or tomorrow)

Thanks Francesco!

Nir