[ovs-dev] RFCv2: OVN database options

Ben Pfaff blp at ovn.org
Fri Mar 11 16:26:25 UTC 2016


On Thu, Mar 10, 2016 at 06:41:51PM -0800, Han Zhou wrote:
> On Thu, Mar 10, 2016 at 5:45 PM, Ben Pfaff <blp at ovn.org> wrote:
> >
> > On Thu, Mar 10, 2016 at 05:31:18PM -0800, Ben Pfaff wrote:
> > > I have been considering this as a minimum interesting scale.  It's hard
> > > for me to know what the interesting scale range is.  I am really happy
> > > to hear what is important to you?
> >
> > That ? was supposed to be an !
> >
> > > Can you tell me about what you want to scale to?
> >
> > But that's really a question.
> 
> It is hard to tell an exact number since it increases over time. But
> considering scale of modern data centers, it is not uncommon to have more
> than 1k hypervisors in a single control plane. Would 5k - 10k clients be a
> realistic target?
> For number of ports, considering number of cores in a BM, maybe something
> around 100 lports per hypervisor sounds better.

That's a significantly higher goal.  10,000 * 100 == 1,000,000 lports,
so if we keep the 1 kB to 5 kB per lport figure then that's 1 GB to 5 GB
of data.

Let me revise my requirements, then:

    - Size: For 1,000 hypervisors at 20 lports/HV and 1 kB/lport, 20 MB;
      for 10,000 hypervisors at 100 lports/HV and 5 kB/lport, 5 GB.

    - Scale: The northbound database has only a single-digit number of
      clients.  Each hypervisor is a client to the southbound
      database, so about 1,000 clients for 1,000 hypervisors or 10,000
      clients for 10,000 hypervisors.

and the analysis:

    - OVSDB.  If we choose to use OVSDB, we'll have to add
      high-availability support.  Also, the table doesn't mention
      scaling, since it's hard to compare objectively, but the OVSDB
      server probably does not scale to 10,000 clients.

(Incidentally, Martin also mentioned today in a meeting that 1,000 HVs
sounded low.)



More information about the dev mailing list