如何通过MobaXterm从远程服务器本地查看TensorBoard？及无传统SSH权限下基于SLURM脚本的实现方案咨询

Hey, great questions—let's tackle them one by one with practical, step-by-step solutions:

Getting TensorBoard from your remote server to your local browser is straightforward with MobaXterm's port forwarding. Here's how:

• First, make sure you can start TensorBoard on the remote server. If you haven't launched it yet, you'll run something like this once connected:

  tensorboard --logdir=/path/to/your/training/logs --port=6006
  

  (You can use any unused port instead of 6006 if needed.)
• Open MobaXterm and set up a new SSH session to your remote server. Before connecting, head to Advanced SSH settings:
  • Enable "Local port forwarding"
  • For Source port, enter a local port (e.g., 6006—matching the remote TensorBoard port makes it easy to remember)
  • For Destination, input localhost:6006 (this points to the TensorBoard instance running on the remote server's localhost)
• Connect to the remote server via this configured SSH session. If you haven't started TensorBoard yet, run the command from the first step now.
• Finally, open your local browser and navigate to http://localhost:6006 (or whatever source port you chose). You'll see your remote TensorBoard dashboard right there!

Since traditional SSH is blocked, we can leverage SLURM's job scheduling to set up a way to access TensorBoard. Here are two reliable approaches:

Option 1: Interactive SLURM Session + Port Forwarding #

If your cluster allows interactive SLURM jobs and limited port forwarding through the login node:

• Submit an interactive job to get a shell on a compute node:

  srun --pty --nodes=1 --ntasks=1 --mem=4G bash
  

  (Adjust resources like --mem to match your cluster's requirements.)
• On the compute node, start TensorBoard with the --bind_all flag (critical to let it listen on all the node's IPs, not just localhost):

  tensorboard --logdir=/path/to/logs --port=6006 --bind_all
  

• From your local machine, use SSH port forwarding through the login node to connect to the compute node's TensorBoard port:

  ssh -L 8080:<compute-node-ip>:6006 your-username@login-node-ip
  

  Replace <compute-node-ip> with the IP address of the compute node you're on (you can get this with hostname -I on the node).
• Now open your browser to http://localhost:8080 to access TensorBoard.

Option 2: Batch SLURM Script with Reverse Port Forwarding #

If interactive sessions aren't feasible, use a batch script to set up a reverse tunnel to the login node:

• Create a SLURM script named tensorboard_slurm.sh:

  #!/bin/bash
  #SBATCH --job-name=tensorboard
  #SBATCH --nodes=1
  #SBATCH --ntasks=1
  #SBATCH --time=24:00:00  # Adjust runtime as needed
  
  # Start TensorBoard, bind to all interfaces
  tensorboard --logdir=/path/to/your/logs --port=6006 --bind_all &
  TENSORBOARD_PID=$!
  
  # Set up reverse port forwarding: forward compute node's 6006 to login node's 8080
  ssh -R 8080:localhost:6006 your-username@login-node-ip -N &
  SSH_TUNNEL_PID=$!
  
  # Keep the job running until either process exits
  wait $TENSORBOARD_PID $SSH_TUNNEL_PID
  

• Submit the script with sbatch tensorboard_slurm.sh
• Once the job starts, you can access TensorBoard by pointing your local browser to http://login-node-ip:8080 (assuming your institution allows access to the login node's ports).

Key Notes: #

• Always check with your cluster admin to confirm which port forwarding methods are allowed (some clusters restrict reverse tunnels).
• Pick unused ports to avoid conflicts—use netstat -tulpn on nodes to check available ports.

内容的提问来源于stack exchange，提问作者stanatomist