[Q&A] How to setup NVFLARE when all clients and server share the same public IP with different ports?

[free-dino]

Python version (python3 -V)

3.12

NVFlare version (python3 -m pip list | grep "nvflare")

2.7.0

NVFlare branch (if running examples, please use the branch that corresponds to the NVFlare version, git branch)

No response

Operating system

Debian 13

Have you successfully run any of the following examples?

• hello-numpy-sag with simulator
• hello-pt with simulator
• hello-numpy-sag with POC
• hello-pt with POC

Please describe your question

I'm trying to run NVFlare in production mode and I'm unsure how to correctly handle provisioning with multiple machines sharing the same public IP.

I have a total of 4 workstations:

• Workstations A and B share the same public IP, but each will run as a client, exposed on a different port.

• Workstations C and D also share another public IP. One will run as a client, and the other will run as the server, also on different ports.

My main issue is how to correctly do the provisioning for this setup. NVFLARE requires specifying hosts in the provisioning config, but the /etc/hosts file does not accept ports, only hostname-to-IP mappings.

My questions:

1. How should I structure project.yml during provisioning when multiple clients and the server share the same public IP but run on different ports?
2. How does NVFlare expect me to handle the /etc/hosts mappings when port numbers can't be included?
3. Can this setup created a potential bottleneck in communicating?

Any example configuration or guidance would be extremely helpful.

[aonexyz]

When multiple NVFLARE clients and the server share the same public IP, provisioning still works correctly as long as every participant uses a unique port. NVFLARE does not require unique IPs only unique host:port combinations.

Here is how to set it up:

1. /etc/hosts does not include ports

This is normal and expected. /etc/hosts should contain only hostname → IP mappings, for example:

/etc/hosts

203.xx.xx.10   clientA
203.xx.xx.10   clientB
203.xx.xx.22   server
203.xx.xx.22   clientC

All machines can use the same IP. Ports are not placed here.

2. project.yml must specify unique ports for each site

Each participant (server or client) gets its own port inside the NVFLARE provisioning config:
participants:

participants:

- name: server
   type: server
   host: server
   components:
     resource_manager:
       ports:
         fed_server: 9002

 - name: clientA
   type: client
   host: clientA
   components:
     resource_manager:
       ports:
         fed_client: 8801

 - name: clientB
   type: client
   host: clientB
   components:
     resource_manager:
       ports:
         fed_client: 8802

 - name: clientC
   type: client
   host: clientC
   components:
     resource_manager:
       ports:
         fed_client: 8803

Even if clientA and clientB share the same public IP, NVFLARE connects to them separately using their assigned ports.

3. Performance / bottlenecks

Sharing the same public IP is not a problem for NVFLARE.
Only bandwidth matters. As long as your upstream speed is sufficient, NVFLARE handles multiple clients behind the same IP without issues.