Open Stack

Report on progress so far:

I was able to fix policies (nova/neutron) to allow tennants to plug to 
'own' external networks, found and report few bugs about error messaging 
in ML2, got working dhcp-agent (on external network! haha). Right now it 
works with cirros (even metadata is ok), but does not work with other 
cloud images, because some issues with metadata. I will continue to 
debug it. It was done with flat networking for single tenant 
(proof-of-concept), but I hope vlan/vxlan stuff is possible to for 
multiple tenants.

Up to now it looks really great from infrastructure (ops) point of view: 
no SPoF on networks nodes, no NAT, no confusing local addresses with 
floatings. Each instance receive white IP and use it as it is. All 
routing is offloaded to network hardware. Cool oldstyle robust 
networking, less non-computing (not paying) hosts, less moving parts. It 
may be not so great from point of view of 'bleeding edge' part of 
openstack (heat/LB/etc), but I do not want my infrastructure to bleed...

I love that thing and I to stick to it.

Updates will follow.

On 12/22/2014 11:35 PM, George Shuklin wrote:
> Thank you for everyone!
>
> After some lurking around I found rather unusual way: use external 
> networks on per-tennant based with directly attached interfaces. This 
> will not only eliminate neutron nodes (as heavy server), but will 
> remove NAT and simplify everything for tenant. All we need just a some 
> VLAN/VXLANs with few external networks (per tenant).
>
> Tenants will have no 'routers' and 'floatingips', but still will have 
> DHCP and other yummy neutron things like private networks with 
> overlapping numbering plans.
>
> Future reports follow.
>
> On 12/21/2014 12:16 AM, George Shuklin wrote:
>> Hello.
>>
>> I've suddenly got request for small installation of openstack (about 
>> 3-5 computes).
>>
>> They need almost nothing (just a management panel to span simple 
>> instances, few friendly tennants), and I curious, is nova-network 
>> good solution for this? They don't want network node and do 'network 
>> node on compute' is kinda sad.
>>
>> (And one more: did anyone tried to put management stuff on compute 
>> node in mild production?)
>