Nice writeup! I like where this is going but see my comments inline below.
On Mon, Nov 26, 2012 at 03:18:22PM -0500, Alon Bar-Lev wrote:
----- Original Message -----
> From: "Livnat Peer" <lpeer(a)redhat.com>
> To: "Shu Ming" <shuming(a)linux.vnet.ibm.com>
> Cc: "Alon Bar-Lev" <abarlev(a)redhat.com>, "VDSM Project
Development" <vdsm-devel(a)lists.fedorahosted.org>
> Sent: Monday, November 26, 2012 2:57:19 PM
> Subject: Re: [vdsm] Future of Vdsm network configuration - Thread mid-summary
>
> On 26/11/12 03:15, Shu Ming wrote:
> > Livnat,
> >
> > Thanks for your summary. I got comments below.
> >
> > 2012-11-25 18:53, Livnat Peer:
> >> Hi All,
> >> We have been discussing $subject for a while and I'd like to
> >> summarized
> >> what we agreed and disagreed on thus far.
> >>
> >> The way I see it there are two related discussions:
> >>
> >>
> >> 1. Getting VDSM networking stack to be distribution agnostic.
> >> - We are all in agreement that VDSM API should be generic enough
> >> to
> >> incorporate multiple implementation. (discussed on this thread:
> >> Alon's
> >> suggestion, Mark's patch for adding support for netcf etc.)
> >>
> >> - We would like to maintain at least one implementation as the
> >> working/up-to-date implementation for our users, this
> >> implementation
> >> should be distribution agnostic (as we all acknowledge this is an
> >> important goal for VDSM).
> >> I also think that with the agreement of this community we can
> >> choose to
> >> change our focus, from time to time, from one implementation to
> >> another
> >> as we see fit (today it can be OVS+netcf and in a few months we'll
> >> use
> >> the quantum based implementation if we agree it is better)
> >>
> >> 2. The second discussion is about persisting the network
> >> configuration
> >> on the host vs. dynamically retrieving it from a centralized
> >> location
> >> like the engine. Danken raised a concern that even if going with
> >> the
> >> dynamic approach the host should persist the management network
> >> configuration.
> >
> > About dynamical retrieving from a centralized location, when will
> > the
> > retrieving start? Just in the very early stage of host booting
> > before
> > network functions? Or after the host startup and in the normal
> > running
> > state of the host? Before retrieving the configuration, how does
> > the
> > host network connecting to the engine? I think we need a basic well
> > known network between hosts and the engine first. Then after the
> > retrieving, hosts should reconfigure the network for later
> > management.
> > However, the timing to retrieve and reconfigure are challenging.
> >
>
> We did not discuss the dynamic approach in details on the list so far
> and I think this is a good opportunity to start this discussion...
>
> From what was discussed previously I can say that the need for a well
> known network was raised by danken, it was referred to as the
> management
> network, this network would be used for pulling the full host network
> configuration from the centralized location, at this point the
> engine.
>
> About the timing for retrieving the configuration, there are several
> approaches. One of them was described by Alon, and I think he'll join
> this discussion and maybe put it in his own words, but the idea was
> to
> 'keep' the network synchronized at all times. When the host have
> communication channel to the engine and the engine detects there is a
> mismatch in the host configuration, the engine initiates 'apply
> network
> configuration' action on the host.
>
> Using this approach we'll have a single path of code to maintain and
> that would reduce code complexity and bugs - That's quoting Alon Bar
> Lev
> (Alon I hope I did not twisted your words/idea).
>
> On the other hand the above approach makes local tweaks on the host
> (done manually by the administrator) much harder.
>
> Any other approaches ?
>
> I'd like to add a more general question to the discussion what are
> the
> advantages of taking the dynamic approach?
> So far I collected two reasons:
>
> -It is a 'cleaner' design, removes complexity on VDSM code, easier to
> maintain going forward, and less bug prone (I agree with that one, as
> long as we keep the retrieving configuration mechanism/algorithm
> simple).
>
> -It adheres to the idea of having a stateless hypervisor - some more
> input on this point would be appreciated
>
> Any other advantages?
>
> discussing the benefits of having the persisted
>
> Livnat
>
Sorry for the delay. Some more expansion.
ASSUMPTION
After boot a host running vdsm is able to receive communication from engine.
This means that host has legitimate layer 2 configuration and layer 3
configuration for the interface used to communicate to engine.
MISSION
Reduce complexity of implementation, so that only one algorithm is used in
order to reach to operative state as far as networking is concerned.
(Storage is extremely similar I can s/network/storage/ and still be relevant).
DESIGN FOCAL POINT
Host running vdsm is a complete slave of its master, will it be ovirt-engine
or other engine.
I do not agree with this direction. It reinforces the single point of failure
of the centralized manager. Also, I am actively working to make vdsm a self
contained component that is independently useful. This proposal will
effectively cripple that effort.
I would prefer a statement that a node _CAN_ be a slave to engine but can also
re-apply a previous configuration in the absense of a management server. See
my other post for how this can be achieved without adding much complexity to the
design.
Having a complete slave ease implementation:
1. Master always apply the setting as-is.
2. No need to consider slave state.
3. No need to implement AI to reach from unknown state X to known state Y + delta.
These would be properties of any intelligent design regardless if engine is
responsible for triggering the configuration changes or if vdsm does it
autonomously. In either case you need to write an algorithm that is capable of
deleting all networking config (except for the management interface). Without
this, you would be unable to apply incremental configuration changes from engine
reliably.
4. After reboot (or fence) host is always in known state.
Once you have a method to strip networking config down to only the management
interface, you can always get back to a known state. I suggest having a vdsm
API that can do this for you.
ALGORITHM
A. Given communication to vdsm, construct required vlan, bonding, bridge setup on
machine.
B. Reboot/Fence - host is reset, apply A.
C. Network configuration is changed at engine:
(1) Drop all resources that are not used by active VMs.
This is the 'network reset' operation I am referring to.
(2) Apply A.
D. Host in maintenance - network configuration can be changed, will
be applied
when host go into active, apply C (no resources are used by VMs, all resources
are dropped).
E. Critical network is down (Host not operative) - network configuration is
not changed.
F. Host unreachable (None responsive) - network configuration cannot be
changed.
BENEFITS
Single deterministic algorithm to apply network configuration.
Pre-defined state after host reboot/fence, host always reachable, previous
network configuration that may be malformed is not in effect.
Do you plan to keep the transactional nature of the current API (ie.
setSafeNetworkingConfig must be called after setupNetworks in order to persist
it)?
Easy to integrate with various network management solution, can it be
primitive iproute, brctl implementation, NetworkManager, OVS or any other
configuration, as Linux is Linux is Linux, the ability to interact with the
kernel is single, while in order to persist implementation requires to
interact with the distribution.
Moreover, a stateless implementation may be integrated with larger set of
network management tools, as no assumption of persistence is added to the
requirements, so if OVS is non-persistent, we use it as-is.
We should aspire to reach to a state in which ovirt-node or any other similar
solution is totally stateless, adding a new node to a cluster should be some
blade rebooting from PXE, each persistence layer we drop, the closer we reach
to managing a large data center built on huge number of machines go up/down as
required joining different clusters.
While discussing clusters, we should also consider autonomic clusters that
enforces policy even if ovirt-engine is unreachable, in this mode we would
like a primitive manager to be able to enforce policy including networking,
while allowing adding/removing nodes without performing any local
configuration.
This could also be done without requiring another redundant management entity by
storing a fallback config to apply when engine is unreachable. Yes, it's
stateful, but that's not always a problem.
IMPLICATIONS
System administrator will not be allowed to modify 'by hand' any of the
network settings (except of this basic engine reachability).
Good luck making that requirement stick in the face of real customers :) You'll
need (at the very least) a hooking mechanism for admins to override some
configuration that hasn't yet been modeled by oVirt.
Special settings can be set in the master, which will apply them via
the
master->vdsm protocol, which in turn use the network management interface in
order to push them, this method should be generic enough to allow pushing most
of the configuration setting allowed (key=value). This approach will also help
replacing/adding nodes in cluster and/or mass deployment.
Edge conditions can be handled by executing some script on host machine,
allowing administrator to override network configuration upon network
configuration event.
SUMMARY
Assuming the host running vdsm as a complete slave and stateless will enable
us to provide better control over that host in the short and long run.
Manual intervention on hosts serving as hypervisors has the flexibility
argument. However at mass deployment, large data-center or dynamic environment
this flexibility argument becomes liability.
Today oVirt plays in the small data center realm so I do think it's important to
give appropriate weight to the flexibility argument. It should be possible to
build different environments based on the needs of the deployment.
--
Adam Litke <agl(a)us.ibm.com>
IBM Linux Technology Center