[Openstack] Making Nova HA summit notes

Vishvananda Ishaya vishvananda at gmail.com
Fri Jul 22 17:10:59 UTC 2011


Below

On Jul 22, 2011, at 5:02 AM, Salvatore Orlando wrote:

> Hi Vish,
>  
> Thanks a lot for this very clear and simple explanation of the ha-networking option.
> At a first glance, it seems this feature should work without any modification even on XenServer deployments. I’ll give it a try and I will let you know!
>  
> It seems that the approach for attributing IPs via DHCP already used in FlatDHCP and VLAN modes also ensures dnsmasq instances running on network hosts do not try to assign the same IP to distinct VMs.
> From your post on unchainyourbrain.com  I understand in multi-host mode the gateway for each VM will be the host where the VM is running. For instance, with reference to the last diagram on the post, VMs on the leftmost host will gateway through 10.0.0.2, whereas VMs on the rightmost one will gateway through 10.0.0.4. If this is correct, VMs on distinct hosts should get different gateways; howeverget_instance_nw_info in nova.network.Manager.NetworkManager sets the gateway for the instance to network[‘gateway’] whereas I was expecting it to be set  to dhcp_server. Am I getting something wrong?

Hmm, interesting point.  In all of the modes that use dnsmasq, the gateway is set as part of the dhcp request, so the mapping isn't used.  It only uses gateway in the mapping for injected modes.  Since the primary use case of the injected modes is to use an external gateway, I think this is ok.  Perhaps we could add a flag that would tell it to use the network gateway vs the dhcp_server ip in the mapping as the gateway.

>  
> Also in the diagram I see the value of the flag flat_network_dhcp_start is 10.0.0.2. Shouldn’t it be 10.0.0.5 (or in general 10.0.0.{#nw_hosts+1})?

The host ips are from the same range as the vm ips.  It is possible for a vm to get an ip 'lower' than a host ip, so it needs to start listening for ips that low.  It will not hand out ips that are unconfigured, so it doesn't cause any problems to start there.
>  
> Thanks again for this nice piece of work (and for mentioning Netstack projects in the “future of networking”)!
>  
> Salvatore
>  
> From: openstack-bounces+salvatore.orlando=eu.citrix.com at lists.launchpad.net [mailto:openstack-bounces+salvatore.orlando=eu.citrix.com at lists.launchpad.net] On Behalf Of Vishvananda Ishaya
> Sent: 21 July 2011 17:24
> To: Mike Scherbakov
> Cc: openstack at lists.launchpad.net
> Subject: Re: [Openstack] Making Nova HA summit notes
>  
> We just recently merged a new HA networking option.  See details in my blog post here:
>  
> http://unchainyourbrain.com/openstack/13-networking-in-nova
>  
> Vish
>  
> On Jul 20, 2011, at 10:04 PM, Mike Scherbakov wrote:
> 
> 
> Hi, 
> Thank you for the work on making nova components HA.
>  
> Did you have a chance to move further in this topic?
> I especially interested in making nova-network HA and looking for possible active-active implementations,
> so the downtime of the service would me minimal.
>  
> Thank you,
>  
> On Tue, May 3, 2011 at 1:22 PM, Edward Konetzko <konetzed at quixoticagony.com> wrote:
> I have attached the slides and Tushar Patil doc on making nova-network ha along with the etherpad notes on the bottom.
> 
> 
> I hope to follow this email up later on in the week with plans for a full reference document based on Cacti.  Thanks for everyone’s participation at the Summit.
> 
> Thanks
> Edward Konetzko
> 
> Etherpad notes
> 
> 
> This Etherpad is for the
> Discussion on Design & Software Considerations for Making Nova HA/Fault Tolerant
> Please put ideas or comments in the appropriate sections
> 
> 
> Database
> - Does zones alleviate the need for HAing the DB?
> 
> 
> 
> RabbitMQ
> For comparison http://wiki.secondlife.com/wiki/Message_Queue_Evaluation_Notes
> - Need to update managers to create persistant queues and messages
> - XMPP an alternate?
> Talk to RabbitMQ devs about
> - Long term can we use Burrow?
> 
> 
> Nova-Network
> NTT Data documentation mailed to openstack list for their heartbeat POC tests
> Are there issues running multiple network nodes and assigning the same IP to mutlple instances?
> How about VRRP protocol?
>   --> we (NTT) are planning to evaluate VRRP using keepalived or some other software. Does anyone knows suitable software?
> 
> 
> Nova-scheduler
> Vish said you can run more then one
> - Yeah with zones and how the scheduler is structured now, it can
> 
> 
> Nova-api
> Possibliy to run this behind real web server, apache, nginx
> 
> 
> Nova-volume
> 
> Nova-Objectstore
> 
> Nova-Compute
> 
> 
> Other ideas
> Services should use dns srv records or something to automate service discovery, this would make running large infrastructures and ipv6 configureation alot easier.
> - zeroconf?  -<-- like the idea but anyone can announce anything in zeroconf it has no idea of a master for security.
> Agreed
> Look at vrrp and keepalived
> 
> 
> Take aways
> Start Discussion with Rabbitmq
> Message Bus needs more investigation
> Discussion on how do we make messages have delievery
> Give feed back to end user, fail or pass just dont leave state in pending forever
> 
> _______________________________________________
> Mailing list: https://launchpad.net/~openstack
> Post to     : openstack at lists.launchpad.net
> Unsubscribe : https://launchpad.net/~openstack
> More help   : https://help.launchpad.net/ListHelp
> 
> 
> 
> 
> -- 
> Mike Scherbakov
> _______________________________________________
> Mailing list: https://launchpad.net/~openstack
> Post to     : openstack at lists.launchpad.net
> Unsubscribe : https://launchpad.net/~openstack
> More help   : https://help.launchpad.net/ListHelp
>  

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack/attachments/20110722/4496930b/attachment.html>


More information about the Openstack mailing list