Multi-Client Setup

Overview

Multi-client setup enables you to deploy federated learning across multiple datasites (data owners) with a central aggregator (data scientist). This guide covers configuration, deployment, and management of distributed FL systems.

Architecture

Components

Aggregator

Data scientist running the FL server

Datasites

Data owners with private datasets

Transport

Communication layer (SyftBox or P2P)

Communication Flow

Setup Process

Step 1: Bootstrap Project

Configure your FL project with aggregator and datasites:

syft_flwr bootstrap /path/to/flower-project \
  --aggregator [email protected] \
  --datasites [email protected],[email protected],[email protected]

This creates a pyproject.toml with:

[tool.syft_flwr]
app_name = "[email protected]_my-fl-project_1234567890"
aggregator = "[email protected]"
datasites = [
    "[email protected]",
    "[email protected]",
    "[email protected]"
]
transport = "syftbox"

Step 2: Distribute Project

Each participant needs the project code:

# Package the project
tar -czf my-fl-project.tar.gz my-fl-project/

# Distribute to all participants
scp my-fl-project.tar.gz [email protected]:/projects/
scp my-fl-project.tar.gz [email protected]:/projects/
scp my-fl-project.tar.gz [email protected]:/projects/

Step 3: Install Dependencies

Each participant installs the project:

# On each datasite
cd /projects/my-fl-project
pip install -e .

Running the Aggregator

Start the Server

On the aggregator machine:

cd /path/to/my-fl-project
python main.py --server

Or explicitly:

python main.py -s

Server Implementation

# run.py:48-73
def syftbox_run_flwr_server(flower_project_dir: Path) -> None:
    pyproject_conf = load_flwr_pyproject(flower_project_dir)
    datasites = pyproject_conf["tool"]["syft_flwr"]["datasites"]
    server_ref = pyproject_conf["tool"]["flwr"]["app"]["components"]["serverapp"]
    app_name = pyproject_conf["tool"]["syft_flwr"]["app_name"]
    
    context = Context(
        run_id=uuid4().int,
        node_id=uuid4().int,
        node_config=pyproject_conf["tool"]["flwr"]["app"]["config"],
        state=RecordDict(),
        run_config=pyproject_conf["tool"]["flwr"]["app"]["config"],
    )
    
    server_app = load_app(
        server_ref,
        LoadServerAppError,
        flower_project_dir,
    )
    
    syftbox_flwr_server(
        server_app=server_app,
        context=context,
        datasites=datasites,
        app_name=app_name,
        project_dir=flower_project_dir,
    )

Server Logs

# Monitor server activity
tail -f server.log

# Expected output:
# INFO: Started SyftBox Flower Server on: [email protected]
# INFO: syft_flwr app name: [email protected]_my-fl-project_1234567890
# INFO: Waiting for nodes to connect...
# INFO: Sampled 3 nodes (out of 3)
# INFO: Received 3/3 results

Running Clients

Start Clients

On each datasite:

cd /path/to/my-fl-project
python main.py  # Default: runs as client

Client Implementation

# run.py:22-45
def syftbox_run_flwr_client(flower_project_dir: Path) -> None:
    pyproject_conf = load_flwr_pyproject(flower_project_dir)
    client_ref = pyproject_conf["tool"]["flwr"]["app"]["components"]["clientapp"]
    app_name = pyproject_conf["tool"]["syft_flwr"]["app_name"]
    
    context = Context(
        run_id=uuid4().int,
        node_id=uuid4().int,
        node_config=pyproject_conf["tool"]["flwr"]["app"]["config"],
        state=RecordDict(),
        run_config=pyproject_conf["tool"]["flwr"]["app"]["config"],
    )
    
    client_app = load_app(
        client_ref,
        LoadClientAppError,
        flower_project_dir,
    )
    
    syftbox_flwr_client(
        client_app=client_app,
        context=context,
        app_name=app_name,
        project_dir=flower_project_dir,
    )

Client Logs

# Monitor client activity
tail -f client.log

# Expected output:
# INFO: Started SyftBox Flower Client on: [email protected]
# INFO: syft_flwr app name: [email protected]_my-fl-project_1234567890
# INFO: 📥 ENCRYPTED Received request, size: 2.34 MB
# INFO: 🔓 Successfully decrypted message
# INFO: Processing message with metadata: ...
# INFO: 🔒 Preparing ENCRYPTED reply, size: 1.87 MB

Data Management

Client Data Structure

Each client should organize their data:

# On [email protected]
/data/
└── hospital1/
    ├── train.csv
    ├── test.csv
    └── metadata.json

# On [email protected]
/data/
└── hospital2/
    ├── train.csv
    ├── test.csv
    └── metadata.json

Loading Client Data

# client_app.py
from pathlib import Path
import pandas as pd
from syft_flwr.utils import run_syft_flwr, get_syftbox_dataset_path

def load_data():
    """Load data for this client."""
    if not run_syft_flwr():
        # Running in standard Flower mode
        return load_flwr_data(partition_id, num_partitions)
    else:
        # Running with Syft-Flwr
        data_dir = get_syftbox_dataset_path()
        
        df_train = pd.read_csv(data_dir / "train.csv")
        df_test = pd.read_csv(data_dir / "test.csv")
        
        return pd.concat([df_train, df_test], ignore_index=True)

Setting Data Path

Environment Variable
Systemd Service
Docker

# On each client machine
export DATA_DIR=/data/hospital1
python main.py

# /etc/systemd/system/fl-client.service
[Unit]
Description=FL Client for Hospital 1
After=network.target

[Service]
Type=simple
User=flclient
WorkingDirectory=/projects/my-fl-project
Environment="DATA_DIR=/data/hospital1"
ExecStart=/usr/bin/python main.py
Restart=always

[Install]
WantedBy=multi-user.target

sudo systemctl enable fl-client
sudo systemctl start fl-client

# docker-compose.yml
version: '3.8'
services:
  fl-client:
    image: my-fl-project:latest
    environment:
      - DATA_DIR=/data
    volumes:
      - /data/hospital1:/data:ro
    command: python main.py
    restart: unless-stopped

docker-compose up -d

Encryption and Security

Key Bootstrap

For SyftBox transport, encryption keys are automatically generated:

# On each participant machine
~/.syftbox/
├── {email}_private_key.pem  # Private encryption key
└── {email}_config.json       # SyftBox config

~/datasites/{email}/
└── did.json                  # Public DID document

Verifying Encryption

Check that encryption is enabled:

# In client/server logs:
# ✅ Expected (encrypted):
INFO: 🔐 End-to-end encryption is ENABLED for FL messages
INFO: 🔒 Preparing ENCRYPTED reply, size: 2.34 MB
INFO: 🔓 Successfully decrypted message

# ⚠️ Warning (not encrypted):
WARNING: ⚠️ End-to-end encryption is DISABLED for FL messages

DID Document Access

Participants must be able to read each other’s DID documents:

# Server can read client DIDs
ls ~/datasites/[email protected]/did.json
ls ~/datasites/[email protected]/did.json

# Clients can read server DID
ls ~/datasites/[email protected]/did.json

Monitoring and Management

Checking Connected Clients

# server_app.py
from flwr.server import ServerApp, Grid
from flwr.common import Context

app = ServerApp()

@app.main()
def main(grid: Grid, context: Context) -> None:
    # Get all connected clients
    all_node_ids = list(grid.get_node_ids())
    logger.info(f"Connected clients: {len(all_node_ids)}")
    
    # Wait for minimum number of clients
    min_clients = 2
    while len(all_node_ids) < min_clients:
        logger.info(f"Waiting for clients... ({len(all_node_ids)}/{min_clients})")
        time.sleep(5)
        all_node_ids = list(grid.get_node_ids())

Health Checks

# Check if a client is responsive
def check_client_health(grid: Grid, node_id: int) -> bool:
    """Send a ping message to check if client is alive."""
    try:
        ping_message = Message(
            content=RecordDict({"action": "ping"}),
            message_type=MessageType.SYSTEM,
            dst_node_id=node_id,
            group_id="health_check",
        )
        
        replies = grid.send_and_receive([ping_message], timeout=30.0)
        return len(replies) > 0 and not replies[0].has_error()
    except Exception as e:
        logger.error(f"Health check failed for node {node_id}: {e}")
        return False

Graceful Shutdown

Stop all clients from the server:

# fl_orchestrator/flower_server.py:50-51
finally:
    syft_grid.send_stop_signal(group_id="final", reason="Server stopped")
    logger.info("Sending stop signals to the clients")

Scaling to Many Clients

Sampling Strategies

For large-scale deployments, sample a subset of clients per round:

import random

@app.main()
def main(grid: Grid, context: Context) -> None:
    num_rounds = context.run_config["num-server-rounds"]
    fraction_sample = context.run_config.get("fraction-sample", 0.1)
    
    for server_round in range(num_rounds):
        # Get all available clients
        all_node_ids = list(grid.get_node_ids())
        
        # Sample a fraction
        num_to_sample = max(2, int(len(all_node_ids) * fraction_sample))
        sampled_nodes = random.sample(all_node_ids, num_to_sample)
        
        logger.info(
            f"Round {server_round + 1}: Sampled {len(sampled_nodes)}/{len(all_node_ids)} clients"
        )
        
        # Send tasks only to sampled clients
        messages = create_messages_for_nodes(sampled_nodes, server_round)
        replies = grid.send_and_receive(messages)

Configuration

# pyproject.toml
[tool.flwr.app.config]
num-server-rounds = 100
fraction-sample = 0.2        # Sample 20% of clients per round
min-available-clients = 5    # Need at least 5 clients
min-sample-size = 2          # Sample at least 2 clients

Example: Three-Hospital Deployment

Project Configuration

# Bootstrap
syft_flwr bootstrap ./diabetes-fl \
  --aggregator [email protected] \
  --datasites [email protected],[email protected],[email protected]

Deployment Steps

Prepare Infrastructure

# On [email protected]
cd /projects/diabetes-fl
pip install -e .

# On [email protected]
cd /projects/diabetes-fl
pip install -e .
export DATA_DIR=/data/hospital-a

# On [email protected]
cd /projects/diabetes-fl
pip install -e .
export DATA_DIR=/data/hospital-b

# On [email protected]
cd /projects/diabetes-fl
pip install -e .
export DATA_DIR=/data/hospital-c

Start Clients First

# On each hospital
python main.py &

# Verify client started
tail -f client.log
# INFO: Started SyftBox Flower Client on: [email protected]

Start Server

# On researcher machine
python main.py --server

# Monitor progress
tail -f server.log
# INFO: Waiting for nodes to connect...
# INFO: Sampled 3 nodes (out of 3)

Monitor Execution

# Check all logs
# Server:
tail -f /projects/diabetes-fl/server.log

# Clients:
ssh [email protected] 'tail -f /projects/diabetes-fl/client.log'
ssh [email protected] 'tail -f /projects/diabetes-fl/client.log'
ssh [email protected] 'tail -f /projects/diabetes-fl/client.log'

Troubleshooting

Clients Not Connecting

Symptom: Server shows “Waiting for nodes to connect…” Solutions:

Verify clients are running:
```
ps aux | grep "python main.py"
```
Check client logs for errors:
```
tail -50 client.log
```

Verify SyftBox is syncing:

ls ~/datasites/{aggregator_email}/flwr/{app_name}/

Messages Not Encrypting

Symptom: Logs show “PLAINTEXT” instead of “ENCRYPTED” Solutions:

Check encryption setting:

echo $SYFT_FLWR_ENCRYPTION_ENABLED  # Should be "true" or empty

Verify DID documents exist:
```
ls ~/datasites/*/did.json
```

Re-bootstrap encryption:

rm -rf ~/.syftbox/*_private_key.pem
python main.py  # Will regenerate keys

Client Can’t Load Data

Symptom: “FileNotFoundError: Path .data/ does not exist” Solutions:

Set DATA_DIR:
```
export DATA_DIR=/path/to/client/data
```

Verify data exists:

ls $DATA_DIR/train.csv
ls $DATA_DIR/test.csv

Performance Issues

Symptom: Slow communication or timeouts Solutions:

Reduce message size (use model compression)
Increase timeouts in config
Sample fewer clients per round
Check network connectivity between participants

Production Deployment

Systemd Service Template

# /etc/systemd/system/[email protected]
[Unit]
Description=Federated Learning Client for %i
After=network-online.target syftbox.service
Wants=network-online.target

[Service]
Type=simple
User=flclient
Group=flclient
WorkingDirectory=/projects/fl-project
Environment="DATA_DIR=/data/%i"
Environment="SYFT_FLWR_ENCRYPTION_ENABLED=true"
ExecStart=/usr/bin/python3 main.py
Restart=always
RestartSec=10
StandardOutput=journal
StandardError=journal

[Install]
WantedBy=multi-user.target

# Enable and start
sudo systemctl enable fl-client@hospital1
sudo systemctl start fl-client@hospital1

# Check status
sudo systemctl status fl-client@hospital1
sudo journalctl -u fl-client@hospital1 -f

Next Steps

Run Simulations

Test multi-client setup locally

Offline Training

Handle intermittent client availability

Transport Configuration

Optimize communication layer

Get Started

Core Concepts

Tutorials

Guides

​Overview

​Architecture

​Components

Aggregator

Datasites

Transport

​Communication Flow

​Setup Process

​Step 1: Bootstrap Project

​Step 2: Distribute Project

​Step 3: Install Dependencies

​Running the Aggregator

​Start the Server

​Server Implementation

​Server Logs

​Running Clients

​Start Clients

​Client Implementation

​Client Logs

​Data Management

​Client Data Structure

​Loading Client Data

​Setting Data Path

​Encryption and Security

​Key Bootstrap

​Verifying Encryption

​DID Document Access

​Monitoring and Management

​Checking Connected Clients

​Health Checks

​Graceful Shutdown

​Scaling to Many Clients

​Sampling Strategies

​Configuration

​Example: Three-Hospital Deployment

​Project Configuration

​Deployment Steps

​Troubleshooting

​Clients Not Connecting

​Messages Not Encrypting

​Client Can’t Load Data

​Performance Issues

​Production Deployment

​Systemd Service Template

​Next Steps