[RFE] New dhcp agents scheduler - use all agents for each network

Bug #1910533 reported by Slawek Kaplonski
6
This bug affects 1 person
Affects Status Importance Assigned to Milestone
neutron
Won't Fix
Wishlist
Slawek Kaplonski

Bug Description

Sometimes in multistack spine&leaf topologies, when dhcp agents are run on compute nodes there is a need to assign each network to be hosted by all dhcp agents. Otherwise vms spawn on some of the sites may not get properly configured IP addresses.

Of course this can be done by setting correct number in "dhcp_agents_per_network" config option but this solution may not scale well as it needs to be updated everytime when new nodes/sites are added.

So maybe better option would be to propose new dhcp scheduler class which can be used in such cases.
This new scheduler would simply always schedule all networks to the all dhcp agents which it will find in the deployment.

Revision history for this message
Slawek Kaplonski (slaweq) wrote :

This rfe was discussed on last drivers meeting: http://eavesdrop.openstack.org/meetings/neutron_drivers/2021/neutron_drivers.2021-01-08-14.00.log.html#l-174 and we agreed to approve it.
Interesting proposal from amotoki is to use existing scheduler with dhcp_agents_per_network=all - I will investigate this solution first.

tags: added: rfe-approved
removed: rfe-triaged
Revision history for this message
LIU Yulong (dragon889) wrote :

-1 for this, please take a look at [1], I mentioned this a bit eailer than this proposal. This has obvious scale issue when the cluster has large sets of resources and hosts.

[1] https://review.opendev.org/c/openstack/neutron-specs/+/768588/1/specs/wallaby/distributed_dhcp.rst#52

Revision history for this message
Slawek Kaplonski (slaweq) wrote :

@LIU - I know that this has scale limitations but this is opt-in option for specific spine-leaf topology use case.

Revision history for this message
Harald Jensås (harald-jensas) wrote :

Could the scheduler ensure N-number of DHCP agents on each segment of a L3 network? I.e ensure DHCP is scheduled on atleast N compute nodes connected to each network segment. This should scale better than scheduling on all nodes, and still fix the issue?

Revision history for this message
Slawek Kaplonski (slaweq) wrote :

After longer investigation it seems for me that our existing scheduler can already schedule network to the dhcp agents "per segment". If subnet has segment configured it will filter host candidates per that segment in https://github.com/openstack/neutron/blob/4381f792c8bd8e5d309ed50288dd4a15a3628e0f/neutron/api/rpc/agentnotifiers/dhcp_rpc_agent_api.py#L205
So this RFE doesn't makes sense really and I'm closing it now.

Changed in neutron:
status: New → Won't Fix
To post a comment you must log in.
This report contains Public information  
Everyone can see this information.

Other bug subscribers

Remote bug watches

Bug watches keep track of this bug in other bug trackers.