Parallel SSH

This skill enables parallel SSH operations across multiple hosts for distributed workloads like SGLang/vLLM prefill-decode disaggregation.

When to Use This Skill

• User provides a list of host IPs (internal or external)
• User wants to run commands in parallel across multiple machines
• User needs to start distributed services (prefill nodes, decode nodes)
• User wants to collect logs from multiple machines simultaneously
• User needs to monitor GPU status across a cluster

SSH Configuration

Default Key File

~/.ssh/google_compute_engine

Step 0: Ensure SSH Key Exists (IMPORTANT)

Before running any SSH commands, ALWAYS check if the key file exists. If not, generate it using gcloud:

# Check if key exists
if [ ! -f ~/.ssh/google_compute_engine ]; then
    echo "SSH key not found, generating via gcloud..."
    gcloud compute config-ssh --quiet
fi

This command:

1. Generates ~/.ssh/google_compute_engine (private key) and ~/.ssh/google_compute_engine.pub (public key)
2. Uploads the public key to the GCP project metadata
3. Configures SSH aliases for all instances in the project

SSH Command Format

ssh -i ~/.ssh/google_compute_engine -o StrictHostKeyChecking=accept-new <IP> "<command>"

The -o StrictHostKeyChecking=accept-new option automatically accepts new host keys (safe for first connection).

Parallel Execution Pattern

Step 1: Parse Host List

When user provides hosts like:

• "10.8.0.81 10.8.0.82 10.8.0.83"
• "10.8.0.81, 10.8.0.82, 10.8.0.83"
• Listed in a file

Parse into individual IPs.

Step 2: Launch Background Tasks

For each host, launch SSH command with run_in_background: true:

ssh -i ~/.ssh/google_compute_engine -o StrictHostKeyChecking=accept-new 10.8.0.81 "command"

IMPORTANT: Launch ALL tasks in a SINGLE message with multiple Bash tool calls to achieve true parallelism.

Step 3: Collect Output

Use TaskOutput to wait for completion, or Read to check progress:

• Output files: /tmp/claude-*/tasks/<task_id>.output

Common Commands

GPU Status Check

nvidia-smi --query-gpu=index,name,memory.used,memory.total,utilization.gpu --format=csv

System Status

vmstat 2 15

Service Status

systemctl status <service> || pgrep -a <process>

Start SGLang Prefill Node

source /opt/deepep/unified-env.sh && \
python3 -m sglang.launch_server \
    --model-path deepseek-ai/DeepSeek-V3 \
    --disaggregation-mode prefill \
    --tp-size 8 \
    --port 30000 \
    --host 0.0.0.0 \
    ...

Start SGLang Decode Node

source /opt/deepep/unified-env.sh && \
python3 -m sglang.launch_server \
    --model-path deepseek-ai/DeepSeek-V3 \
    --disaggregation-mode decode \
    --tp-size 8 \
    --port 30001 \
    --host 0.0.0.0 \
    ...

Example Workflows

Example 1: Check GPU Status Across Cluster

User: "Check GPU status on 10.8.0.81 10.8.0.82 10.8.0.83"

Action: Launch 3 parallel SSH tasks with nvidia-smi command, collect and summarize results.

Example 2: Start Distributed Inference

User: "Start prefill on 10.8.0.81, decode on 10.8.0.82 and 10.8.0.83"

Action:

1. SSH to 10.8.0.81: Start prefill node
2. SSH to 10.8.0.82: Start decode node 1
3. SSH to 10.8.0.83: Start decode node 2 (All launched in parallel)

Example 3: Kill Services Across Cluster

User: "Kill all python processes on all machines"

Action: Launch parallel pkill -f python commands on all hosts.

Example 4: Collect Logs

User: "Get last 100 lines of sglang logs from all machines"

Action: Launch parallel tail -100 /path/to/logs commands.

Host Key Management

If SSH fails with "Host key verification failed":

# Add host key to known_hosts
ssh-keyscan -H <IP> >> ~/.ssh/known_hosts

# Or use StrictHostKeyChecking=accept-new (recommended)
ssh -o StrictHostKeyChecking=accept-new ...

Best Practices

1. Always use background mode for long-running commands
2. Launch all tasks in single message for true parallelism
3. Collect output progressively for long tasks
4. Summarize results in table format for easy comparison
5. Handle failures gracefully - report which hosts failed

Output Format

When reporting results, use table format:

Host	Status	Key Metrics
10.8.0.81	OK	GPU 0-7: 95% util
10.8.0.82	OK	GPU 0-7: 80% util
10.8.0.83	FAILED	SSH timeout

Environment Variables for Distributed Training

Common environment variables to set on each node:

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export NCCL_SOCKET_IFNAME=enp0s19
export GLOO_SOCKET_IFNAME=enp0s19
export MASTER_ADDR=<prefill_node_ip>
export MASTER_PORT=29500

Troubleshooting

SSH Key Not Found

If you see "Identity file not accessible" or "No such file or directory":

# Generate SSH key using gcloud (recommended for GCE)
gcloud compute config-ssh --quiet

# This creates ~/.ssh/google_compute_engine and uploads public key to project metadata

SSH Connection Refused

• Check if SSH service is running on target
• Verify network connectivity: ping <IP>

Permission Denied (publickey)

1. First, check if key exists:

   ls -la ~/.ssh/google_compute_engine
   

2. If not exists, generate it:

   gcloud compute config-ssh --quiet
   

3. If exists but still fails, verify key is uploaded to project:

   gcloud compute project-info describe --format="value(commonInstanceMetadata.items.filter(key:ssh-keys))"
   

Command Timeout

• Use nohup for long-running processes
• Use screen or tmux for persistent sessions
• Increase Bash timeout parameter