[openstack-dev] [Quantum] continuing todays discussion about the l3 agents
Gary Kotton
gkotton at redhat.com
Mon Dec 3 10:19:32 UTC 2012
On 12/03/2012 11:53 AM, gong yong sheng wrote:
> On 12/03/2012 05:32 PM, Gary Kotton wrote:
>> On 12/03/2012 04:16 AM, gong yong sheng wrote:
>>> On 12/03/2012 02:29 AM, Vinay Bannai wrote:
>>>> My understanding of the "scheduler" approach based on what I read
>>>> on the ML's is to have a mechanism where the DHCP agents can land
>>>> on different nodes. For example, just like we have compute hosts in
>>>> nova, we have a bunch of DHCP capable hosts (and L3 capable hosts
>>>> etc) that can be selected to host the network service for a tenant
>>>> when the network/subnet is created. The process of selecting the
>>>> host to run the service is based on a "scheduler". This allows a
>>>> graceful horizontal scaling. This approach is similar to what nova
>>>> does. You have a bunch of hosts capable of providing a network
>>>> service and the "scheduler" picks them based on filters and other
>>>> tunable knobs. I think you already know this:-). I was spelling it
>>>> out so that you can see where I am coming from.
>>> If we don't want all dhcp agents to host the data of all the networks,
>>> My Idea is:
>>> 1. let quantum server have the ability to know all about the dhcp
>>> agents. for example we can have quantum agents-list to show all the
>>> agents running in the quantum deloyment,
>>> and the network they are hosting.
>>
>> ok, this may be useful for debugging. will this display and have the
>> status of the dhcp agent, say for example i deploy and agent and it
>> has an exception do to a bug? nn
> I think it is good for management, not just good for debugging. If the
> agent cannot start, the quantum agents-list will not have it in a :)
> status. If it has exception and is still running, we can have it
> report its status with exception. But regarding the log, I think we
> should leave it to upper management tools. For example, admin user can
> start the agent with integrated log facility.
Visibility is always good.
>>> 2. let admin user have the ability to config the dhcp agents what
>>> networks they should host. For example, quantum dhcpagent-update
>>> dhcpagent1 --networks network1 network2 network3. or quantum
>>> net-create network1 --dhcpagents agent1 agent2. And if admin user
>>> does not specify which agent to host which network, we can let
>>> scheduler to decide automatically
>>
>> this is exactly what i am suggesting, except we do not need to change
>> the quantum api to provide this. the agent can receive this as an
>> input parameter. in principle we agree on what needs to be done, but
>> the question is how.
> I am suggesting to control agents from quantum server. we can use
> quantum cli or API to command which agents host which networks. Of
> course admin user also can configure the dhcp agents to accept only
> some networks data. We can support both.
> We can add quantum api ( maybe by extension) and store the networks
> and dhcp agents mapping in db. then notify related dhcp agents with
> related data.
I personally do not like this approach and I think that the agents
should be independent of the service (as much as possible). At the
moment the DHCP agent gets its configuration via the notifications. I
have a number of questions:
1. do you plan to continue to use the same mechanism of implement direct
communication with the specific agents?
2. just because a compute node/network node loses connectivity with the
quantum service does not mean that it loses network connectivity (and
vice versa). How will we deal with issues like this?
3. Can you please clarify more on the extension. will this be on the
subnet (that is where there is a flag for dhcp status)
>>> So for scale vertically:
>>> we can specify much agents host some same networks
>>> So for scale horizontally:
>>> we can add as many as dhcp agents. quantum scheduler will distribute
>>> new networks automatically or admin user can specify.
>>
>> i have a number of problems with a quantum "scheduler". the first
>> being a single point of failure. the second being the fact that it
>> needs to be aware of the state and load of a dhcp agent. how will the
>> scheduler provide for HA?
> What do you mean by single point of failure? In fact we don't need to
> run it in a separate binary at all. It is just a module of
> quantum-server.
> state and load of a dhcp agent is in db. If you are meaning the
> quantum-server's HA, I don't think we already have a good one by now.
> I think we should break the current quantum-server into two parts:
> 1. one part is for just API (rest): for which we can use multiple
> process just like other nova api servers. and operator will also be
> able to use load balancer to HA it.
> 2. anyone part is for incoming queue message processing. We can run
> multiple this part nodes, this is the AMQP feature to scale out.
>>
>>
>>> For us to run multiple dhcp agents, we need to make sure our dhcp
>>> anti spoofing work.
>>>>
>>>> Either way we look at it, I think it will be helpful if we
>>>> decoupled the horizontal (scaling to multiple nodes) and vertical
>>>> scaling (redundancy and failover). One should not imply the other.
>>>> In your last paragraph, you mention "orchestration tool" and dhcp
>>>> agents configured to handle specific networks. I have not been able
>>>> to wrap my head around this completely but it appears to b ea
>>>> different variant of the "scheduler" approach where it is
>>>> configured manually. Is my understanding correct? Or if you don't
>>>> mind, can you elaborate further on that idea.
>>>>
>>>> Thanks
>>>> Vinay
>>>>
>>>> On Sun, Dec 2, 2012 at 7:16 AM, Gary Kotton <gkotton at redhat.com
>>>> <mailto:gkotton at redhat.com>> wrote:
>>>>
>>>> On 12/01/2012 03:31 AM, gong yong sheng wrote:
>>>>> On 12/01/2012 07:49 AM, Vinay Bannai wrote:
>>>>>> Gary and Mark,
>>>>>>
>>>>>> You brought up the issue of scaling horizontally
>>>>>> and vertically in your earlier email. In the case of
>>>>>> horizontal scaling, I would agree that it would have to be
>>>>>> based on the "scheduler" approach proposed by Gong and Nachi.
>>>>
>>>> I am not sure that I understand the need for a scheduler when
>>>> it comes to the DHCP agent. In my opinion this is unnecessary
>>>> overhead and it is not necessarily required.
>>>>
>>>> Last week Mark addressed the problem with all of the DHCP
>>>> agents all listening on the same message queue. In theory we
>>>> are able to run more than one DHCP agents in parallel. This
>>>> offers HA at the expense of an IP per DHCP agent per subnet.
>>>>
>>>> I think that for the DHCP agents we need to look into enabling
>>>> the DHCP agents to treat specific networks. This can be done in
>>>> a very rudimentary way - have a configuration variable for the
>>>> DHCP agent indicating a list of networks to be treated by the
>>>> agent. A orchestration tool can just configure the network ID's
>>>> and launch the service - then we will have scalable and highly
>>>> available DHCP service. I would prefer not to have to add this
>>>> into the Quantum API as it just complicates things.
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>>> On the issue of vertical scaling (I am using the DHCP
>>>>>> redundancy as an example), I think it would be good to base
>>>>>> our discussions on the various methods that have been
>>>>>> discussed and do pro/con analysis in terms of scale,
>>>>>> performance and other such metrics.
>>>>>>
>>>>>> - Split scope DHCP (two or more servers split the IP address
>>>>>> and there is no overlap)
>>>>>> pros: simple
>>>>>> cons: wastes IP addresses,
>>>>>>
>>>>>> - Active/Standby model (might have run VRRP or hearbeats to
>>>>>> dictate who is active)
>>>>>> pros: load evenly shared
>>>>>> cons: needs shared knowledge of address assignments,
>>>>>> need hearbeats or VRRP to keep track of failovers
>>>>> another one is the IP address waste. we need one VIP, and 2+
>>>>> more address for VRRP servers. ( we can use dhcp server's ip
>>>>> if we don't want to do load balancing behind the VRRP servers)
>>>>> another one is it will make system complicated.
>>>>>>
>>>>>> - LB method (use load balancer to fan out to multiple dhcp
>>>>>> servers)
>>>>>> pros: scales very well
>>>>>> cons: the lb becomes the single point of failure,
>>>>>> the lease assignments needs to be shared between
>>>>>> the dhcp servers
>>>>>>
>>>>> LB method will also wast ip address. First we at lease need a
>>>>> VIP address. then we will need more dhcp servers running for
>>>>> one network.
>>>>> If we need to VRRP the VIP, we will need 2+ more addresses.
>>>>> another one is it will make system complicated.
>>>>>> I see that the DHCP agent and the quantum server communicate
>>>>>> using RPC. Is the plan to leave it alone or migrate it
>>>>>> towards something like AMQP based server in the future when
>>>>>> the "scheduler" stuff is implemented.
>>>>> I am not very clear your point. But current RPC is on AMQP.
>>>>>>
>>>>>> Vinay
>>>>>>
>>>>>>
>>>>>> On Wed, Nov 28, 2012 at 8:03 AM, Mark McClain
>>>>>> <mark.mcclain at dreamhost.com
>>>>>> <mailto:mark.mcclain at dreamhost.com>> wrote:
>>>>>>
>>>>>>
>>>>>> On Nov 28, 2012, at 8:03 AM, gong yong sheng
>>>>>> <gongysh at linux.vnet.ibm.com
>>>>>> <mailto:gongysh at linux.vnet.ibm.com>> wrote:
>>>>>>
>>>>>> > On 11/28/2012 08:11 AM, Mark McClain wrote:
>>>>>> >> On Nov 27, 2012, at 6:33 PM, gong yong sheng
>>>>>> <gongysh at linux.vnet.ibm.com
>>>>>> <mailto:gongysh at linux.vnet.ibm.com>> wrote:
>>>>>> >>
>>>>>> >> Just wanted to clarify two items:
>>>>>> >>
>>>>>> >>>> At the moment all of the dhcp agents receive all of
>>>>>> the updates. I do not see why we need the quantum service
>>>>>> to indicate which agent runs where. This will change the
>>>>>> manner in which the dhcp agents work.
>>>>>> >>> No. currently, we can run only one dhcp agent since
>>>>>> we are using a topic queue for notification.
>>>>>> >> You are correct. There is a bug in the underlying
>>>>>> Oslo RPC implementation that sets the topic and queue
>>>>>> names to be same value. I didn't get a clear explanation
>>>>>> of this problem until today and will have to figure out a
>>>>>> fix to oslo.
>>>>>> >>
>>>>>> >>> And one problem with multiple agents serving the same
>>>>>> ip is:
>>>>>> >>> we will have more than one agents want to update the
>>>>>> ip's leasetime now and than.
>>>>>> >> This is not a problem. The DHCP protocol was designed
>>>>>> for multiple servers on a network. When a client accepts
>>>>>> a lease, the server that offered the accepted lease will
>>>>>> be the only process attempting to update the lease for
>>>>>> that port. The other DHCP instances will not do
>>>>>> anything, so there won't be any chance for a conflict.
>>>>>> Also, when a client renews it sends a unicast message to
>>>>>> that previous DHCP server and so there will only be one
>>>>>> writer in this scenario too. Additionally, we don't have
>>>>>> to worry about conflicting assignments because the dhcp
>>>>>> agents use the same static allocations from the Quantum
>>>>>> database.
>>>>>> > I mean dhcp agent is trying to update leasetime to
>>>>>> quantum server. If we have more than one dhcp agents,
>>>>>> this will cause confusion.
>>>>>> > def update_lease(self, network_id, ip_address,
>>>>>> time_remaining):
>>>>>> > try:
>>>>>> >
>>>>>> self.plugin_rpc.update_lease_expiration(network_id,
>>>>>> ip_address,
>>>>>> >
>>>>>> time_remaining)
>>>>>> > except:
>>>>>> > self.needs_resync = True
>>>>>> > LOG.exception(_('Unable to update lease'))
>>>>>> > I think it is our dhcp agent's defect. Why does our
>>>>>> dhcp agent need the lease time? all the IPs are managed
>>>>>> in our quantum server, there is not need for dynamic ip
>>>>>> management in dhcp server managed by dhcp agent.
>>>>>>
>>>>>> There cannot be confusion. The dhcp client selects only
>>>>>> one server to accept a lease, so only one agent will
>>>>>> update this field at a time. (See RFC2131 section 4.3.2
>>>>>> for protocol specifics). The dnsmasq allocation database
>>>>>> is static in Quantum's setup, so the lease renewal needs
>>>>>> to propagate to the Quantum Server. The Quantum server
>>>>>> then uses the lease time to avoid allocating IP addresses
>>>>>> before the lease has expired. In Quantum, we add an
>>>>>> additional restriction that expired allocations are not
>>>>>> reclaimed until the associated port has been deleted as well.
>>>>>>
>>>>>> mark
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> OpenStack-dev mailing list
>>>>>> OpenStack-dev at lists.openstack.org
>>>>>> <mailto:OpenStack-dev at lists.openstack.org>
>>>>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>> --
>>>>>> Vinay Bannai
>>>>>> Email: vbannai at gmail.com <mailto:vbannai at gmail.com>
>>>>>> Google Voice: 415 938 7576 <tel:415%20938%207576>
>>>>>>
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> OpenStack-dev mailing list
>>>>>> OpenStack-dev at lists.openstack.org <mailto:OpenStack-dev at lists.openstack.org>
>>>>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>>>>
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> OpenStack-dev mailing list
>>>>> OpenStack-dev at lists.openstack.org <mailto:OpenStack-dev at lists.openstack.org>
>>>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>>>
>>>>
>>>> _______________________________________________
>>>> OpenStack-dev mailing list
>>>> OpenStack-dev at lists.openstack.org
>>>> <mailto:OpenStack-dev at lists.openstack.org>
>>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>>>
>>>>
>>>>
>>>>
>>>> --
>>>> Vinay Bannai
>>>> Email: vbannai at gmail.com <mailto:vbannai at gmail.com>
>>>> Google Voice: 415 938 7576
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> OpenStack-dev mailing list
>>>> OpenStack-dev at lists.openstack.org
>>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20121203/18986f57/attachment.html>
More information about the OpenStack-dev
mailing list