6. Configure the Campaign and Start the Server

📍 Where you are: Task Owner 6/7 — ① Design → ② Create Task → ③ Workspace → ④ Train · publish → ⑤ Approve → ⑥ Campaign → ⑦ Results
Screen: FedOps Console · Other side: when you start the server, the Client starts joining in Client Track ch. 4 · Next: watch training and read the results

🎯 What you do in this step: Set the rules of this one run of federated learning (campaign) — how many rounds, how many participants per round, and how to merge — and start the aggregation server that collects and merges parameters. This step has the most options in the tutorial and the most direct effect on the results.

  1. 1) Create the aggregation server
  2. 2) Set the campaign rules
    1. ⚙️ Options explained — campaign
      1. Global model evaluation
  3. 3) Start the server
    1. ⚙️ Explained — what Start FL server fixes
  4. 4) Tell the Client
  5. ✅ Checklist
  6. If something goes wrong
  7. Next
flowchart LR
  A["Create server<br/>(allocate resources)"] --> B["Save campaign<br/>rules"] --> C["Start server<br/>wait for round 1"] --> D["Clients connect<br/>rounds run"]

1) Create the aggregation server

Go to FedOps Console → My Federated Tasks → handson-heart.

Server Management tab

  1. Click the Server Management tab.

Create scalable server button

  1. Scroll down and click Create scalable server under Server lifecycle. It starts creating immediately, with no confirmation dialog.

FL Server created state and Refresh status

Wait until the FL server field in Runtime overview at the top goes from FL Server Creating → FL Server created. Measured 1 min 30 s on 2026-10-20.

  1. If it does not change, click Refresh status.

“Create” and “start” are different. Here you only reserve the place (compute resources) in the cloud for the server to run; federated learning has not started yet.

2) Set the campaign rules

Rounds 2, Clients per round 2, Check evaluation setup, Save campaign

  1. Set Rounds.
  2. Set Clients per round. The default is 1; change it to 2. (See the table below.)
  3. Click Check evaluation setup. Click it after the server is FL Server created; before that you get Validation check unavailable.
  4. Click Save campaign.
SituationRoundsClients per round
Practicing alone (the Owner is also the Client)21
This tutorial (Owner + 1 Client)22 — rounds proceed only when both hospitals join
Accepting more hospitals (e.g. 3 with the VA hospital)2 or moreActual number of participants

Evaluation setup ready and Saved

Done when you see green Evaluation setup ready and Saved.

⚙️ Options explained — campaign

OptionWhat it decidesIf you increase or change itThis tutorial
RoundsHow many times the server loops “send → local training → merge” (1 or more)More can refine the Global Model, but time and communication grow, and at some point it hardly improves2 — short, since the goal is to see the flow. Also the FedOps template’s recommended value
Clients per roundHow many must finish training in a round before the server merges (1 or more)More reflects more data per round, but the server waits until that many gather. If larger than the real number of participants, rounds never start2 — the number who will actually take part (Owner + Client)
Aggregation strategyHow parameters are mergedIn FedOps 1.3 only FedAvg can be chosen. The list comes from the methods the published Release declaresFedAvg
Global model evaluationWhere the Global Model is scoredExplained separately belowValidation OFF · Client metrics

Global model evaluation

ChoiceWhere scoring happensNeedsTrade-offs
Validation OFF · Client metricsEach Client evaluates on its own test data, and the server takes a weighted average proportional to data sizeNothingThe Owner does not need separate validation data. But if hospitals’ data differ, so do the yardsticks
Validation ON · Server evaluationThe server evaluates the Global Model directly on common validation data uploaded by the OwnerServer validation dataset uploaded from Studio, passing Check evaluation setupEvery round is compared with the same yardstick. But the Owner must hold representative validation data, and that data goes to the server

This tutorial uses OFF: it is the basic form where federated learning runs without the Owner uploading data to the server. The template’s config.yaml also has server_evaluation.enabled: false. ON is not covered in this tutorial.

📚 Federated learning background — FedAvg and rounds

FedAvg (Federated Averaging), proposed by Google in 2017, is the most basic aggregation method. It averages the parameters each client sends with weights proportional to the amount of data.

Global = (n₁ × Hungary's parameters + n₂ × Cleveland's parameters) / (n₁ + n₂)
         n = rows each trained on this round (the Client sends it to the server with the parameters)

Hospitals with more data have a bigger say. In this tutorial Cleveland (303 rows) counts slightly more than Hungary (261 rows). Under non-IID, the Global Model leans toward the distribution of the hospital with more data.

Balancing rounds and local training. Within each round a client trains several times on local data. More local training means fewer rounds (less communication) are needed, but under non-IID each client drifts far toward its own data (client drift), and merging can even make things worse. Methods such as FedProx were introduced to reduce this (not available in FedOps 1.3).

Clients per round is also a rule for “waiting”. The server waits until the set number of participants finish training. If a Client leaves with Stop Client, this value does not shrink automatically; the server keeps waiting for other eligible Clients.

3) Start the server

Start FL server and Show live log

  1. Click Start FL server in the Command console.
  2. Click Show live log.

[ROUND 1] in the live log

When [ROUND 1] and server start by round appear at the bottom of the Live server log, the server is waiting for participants.

Scrolling up in the live log shows many AWS-related DEBUG lines. One of them prints the server’s access key ID, so do not capture and share the live log screen. (A defect to report to the FedOps team.)

⚙️ Explained — what Start FL server fixes

Start FL server fixes the following as one bundle.

FixedMeaning
Saved campaign rulesRounds, Clients per round, and aggregation method for this run
New campaign run IDSo this run’s records do not mix with other runs
Published ReleaseSo every Client trains with the same code (chapter 4)
Starting modelThe Initial Model for the first campaign, the previous Global Model vN for later ones

So changing campaign values while the server runs does not affect this run. Starting a duplicate server for the same Task while one is running is also blocked.

ButtonWhat it does
Stop serverStops the running federated learning process
Pause (Server lifecycle)Returns compute resources while keeping the needed data. Click it when the practice is over
ResumeBrings a paused server back

4) Tell the Client

Now, while the server waits at [ROUND 1], is when the Client joins. Tell the Client “the server is up”. The Owner is also a participant (the Hungarian hospital), so the Owner must also click Start Client in Studio as in Client Track ch. 3 and ch. 4 to fill the two seats. The Client follows Client Track ch. 4.

Starting the server requires the readiness check again. When a new campaign server starts, Studio’s Task Data goes back to Check required for both the Owner and the Client. Before Start Client, click Check Participation Readiness again and confirm ✓ Local Ready. It took 1 min 16 s on the Mac (Owner) on 2026-10-20.

If you practice alone, go to Studio now and do Client Track ch. 3 and ch. 4 with your own account. The Owner’s Workspace already has the Hungarian data.

✅ Checklist

  • You saw FL Server created in Server Management
  • Clients per round matches the actual number of participants
  • You saw Evaluation setup ready · Saved
  • [ROUND 1] appeared in the live log
  • You told the Client the server is up

If something goes wrong

SymptomFix
Start FL server is disabledThe server is still Creating. Wait for created with Refresh status
Validation check unavailableYou clicked Check evaluation setup before the server was created. Wait and click again
The server does not start roundsRounds start only when Clients per round Clients connect. Check the approved and actual participants. Check that the Owner also clicked Start Client
Studio keeps showing Server Offline after the server startedStudio’s display can lag several minutes. Click Refresh on the federated learning screen
Start Client fails with [Errno 2] No such file or directory: '…/task-binding.tmp'Probably clicked while Studio was writing a file right after the readiness check (happened once on a Mac on 2026-10-19). Click Refresh and click Start Client again
The connection count in Console Monitoring (online now) does not changeMonitoring refreshes every few minutes. Use Studio’s state (waiting_round, training) instead

Next

7. Watch Training and Read the Results


Back to top

© 2026 Cognitive Computing Lab, Gachon University