Parallel GPU visualization example of trame ?

[nyue]

Hi,

I have been given a brief (1 week) access to a cluster of GPUs with a separate head/jumpbox node

I am wondering if there are existing example I can use to showcase Trame in such a cluster, hopefully using some volume rendering example.

I am happy to roll up my sleeve to modify/change example to adapt but I am hoping to start with a known working example.

Kind regards

[jourdain]

For parallel rendering, you would need to use ParaView 5.13. If you already have a dataset, load it in ParaView, tweak the rendering to your liking, and save the state as a Python file.

Then add the following lines into a trame-app.py file:

import paraview.web.venv
import ptc

# [... python state file content...]

web_app = ptc.Viewer(from_state=True)
web_app.start()

Then on your Cluster, create a venv using the same Python as your ParaView (3.10)

python3.10 -m venv /test/pv-venv
source /test/pv-venv
pip install paraview-trame-components

Then start your MPI job running pvbatch that

export PV_VENV=/test/pv-venv
/paraview/bin/mpi [...] /paraview/bin/pvbatch /path/to/your/trame-app.py

HTH

[nyue]

Thanks for the suggestion. Learning curve too steep to wrangle this into a working example as my time is limited. Ended up choosing to go via the traditional mpiexec+pvserver on multiple hosts to showcase.

If I get an extension, I may revisit.

[jourdain]

You could call paraview.simple.Connect(...) in a trame script to connect to your mpiexec+pvserver.

[nyue]

Ah! That is interesting, I will look into that.

[nyue]

To keep things simple, I am testing locally to get a hang of trame

On my local box, I have /home/nyue/systems/ParaView/5.13.1/ParaView-5.13.1-egl-MPI-Linux-Python3.10-x86_64/bin

I start a session with the following

 ./mpiexec -hosts head0 ./pvserver

I created the following virtual env via conda

name: trame
channels:
  - conda-forge
  - default
dependencies:
  - python=3.10.13
  - paraview=5.13.1
  - trame
  - trame-vtk
  - trame-vuetify

I tried a simple connection with the following script

from paraview.simple import *

# Connect to the pvserver
server = servermanager.Connect("head0", 11111)

# Load your data - replace `GetSources()` with appropriate pipeline creation
data_source = GetSources()  # Or replace with specific data load logic

# Example of loading a built-in example if no data source is present
if not data_source:
    data_source = Wavelet()  # Example built-in data source for testing

I connect via the created conda environment I named trame

(trame) nyue@head0:~/temp/trame/simple$ python mypvserver.py 
Segmentation fault (core dumped)

I am getting the following error

Waiting for client...
Connection URL: cs://head0:11111
Accepting connection(s): head0:11111
Client connected.
(   6.597s) [pvserver        ]vtkSocketCommunicator.c:774    ERR| vtkSocketCommunicator (0x9c45c0): Could not receive tag. 1
(   6.598s) [pvserver        ]vtkTCPNetworkAccessMana:284    ERR| vtkTCPNetworkAccessManager (0x8d3790): Some error in socket processing.
Exiting...

[jourdain]

You can not use the paraview from conda as it is incompatible with your pvserver. You need to use pvpython from the ParaView 5.13.1 that you download from our website.

Then create a venv from a python 3.10 with all the trame dependencies and run it like below

pvpython --venv /path/to/venv /path/to/trame-app.py

[nyue]

Understood.

I have since made some head way, I tested it on my laptop and now trying to deploy to the server

(.venv) ubuntu@san8oaict:~/projects$ pvbatch state-loader.py --server
ParaView is using venv: /home/ubuntu/projects/.venv

App running at:
 - Local:   http://localhost:8080/
 - Network: http://127.0.0.1:8080/

I was wondering if there is a way for the server to listen on host/IP 0.0.0.0 ? like it does with http.server

(.venv) ubuntu@san8oaict:~/projects$ python3 -m http.server
Serving HTTP on 0.0.0.0 port 8000 (http://0.0.0.0:8000/) ...
38.88.120.18 - - [02/Nov/2024 19:26:37] "GET / HTTP/1.0" 200 -
^C
Keyboard interrupt received, exiting.

Thank you.

[nyue]

Answering my own question, I can use --host 0.0.0.0 to achieve that