Skip to content

Speeding up Filters & Rendering with Client Server configurations

For Both of these configurations it is recommended that you start up a remote desktop and launch paraview according to the instructions on the previous documented page Getting Started With ParaView. Using the Embedded graphics library paraview binary

Steps To launch a EGL server

This workflow uses offscreen rendering and a gpu to speed up interactive workflows.

First request an allocation on Ocelote that has 1 gpu Please remember to replace the <your account> with your own account name

salloc -A <your account> -p standard --gres=gpu:1 -t 3:00:00 -N 1 -n 16

The following lines will download a version of paraview that we can use for offscreen headless rendering. The command will actually rename the export file in order to make it easier to uncompress it.

wget "https://www.paraview.org/paraview-downloads/download.php?submit=Download&version=v5.11&type=binary&os=Linux&downloadFile=ParaView-5.11.1-egl-MPI-Linux-Python3.9-x86_64.tar.gz" -O paraview_egl.tar.gz

Here we uncompress the tar file

tar xf paraview_egl.tar.gz

The folder we get from this will probably include the full paraview version name so we will use a * pattern match character to go into the bin folder where the pvserver program we need to use resides

cd ParaView-5.11.1*/bin/

Here's where the exciting part occurs. We will use the pvserver program to start up a process which listens for connections from the graphical interface and fulfills requests on its behalf. The extra benefit is that it will use GPU hardware accelerated rendering to get back results to the viewport even faster. We will also be making use of the embedded graphics library EGL, to perform these renderings without the need for an x-server display. The flag --displays is used to specify the gpu card index that we want to make use of, starting from the number 0. So if we had another card we wished to use instead we would use --displays=1. The potential to leverage multiple gpus is available but requires a different compiled version of paraview with a specific nvidia license so it will not be covered here.

./pvserver --displays=0

For completeness here are the commands all together for copying in pasting into a terminal on the HPC

wget "https://www.paraview.org/paraview-downloads/download.php?submit=Download&version=v5.11&type=binary&os=Linux&downloadFile=ParaView-5.11.1-egl-MPI-Linux-Python3.9-x86_64.tar.gz" -O paraview_egl.tar.gz
cd ParaView-5.11.1*/bin/
tar xf paraview_egl.tar.gz
./pvserver --displays=0

Take note of the address that is written to the screen at this point it will look something like cs://<node_name>.<cluster_name>.arizona.edu:11111 and this will be used when we connect via the gui client Steps to launch a client to connect to the server

For a first draft of this page we will rely on the documentation provided on the paraview read the docs page to understand the process of connecting from the graphical user interface. https://docs.paraview.org/en/latest/ReferenceManual/parallelDataVisualization.html#configuring-a-server-connection

Once the connection has been made, it is possible to determine the extent to which the gpu is utilized by putting the running ./pvserver in the background with ctrl-z and then typing bg. At this point you can now type nvidia-smi -l 2 to get a log of the activity on the gpu in a text table.

Using multinode MPI and osmesa for

This process is very similar to the above, but we will be asking for a slightly different allocation. Use the following command on Elgato for testing having two full nodes of 16 cpu cores to speed up filtering and rendering with paraview. After this tutorial feel free to test out using different machines or node and core counts. The rule being that the total number of cpu cores is specified in the -n flag and the total number of nodes is specified with the -N flag. For more information consult the basic Slurm scheduling pages. Please remember to replace the with your own account name

interactive -a <your account> -N 2 -n 32 -t 1:00:00

When the first of the multiple nodes is allocated go and download the latest paraview osmesa.

wget "https://www.paraview.org/paraview-downloads/download.php?submit=Download&version=v5.11&type=binary&os=Linux&downloadFile=ParaView-5.11.1-osmesa-MPI-Linux-Python3.9-x86_64.tar.gz" -O paraview_mesa.tar.gz

The next step is to untar it

tar xf paraview_mesa.tar.gz

Then at this point you are ready to navigate into the bin folder

cd Paraview*/bin Now use the provided mpiexec binary to start up a pvserver backed up by all our allocated cores

./mpiexec -n 32 ./pvserver

Once again this is an offscreen rendering system able to run without needing to have displays or anything. A side benefit is since this is an entire osmesa build, no missing opengl dependencies cause problems. Again we need to pay attention to the output of this command so that we can see the server's connection url to use. For instance if the mpi nodes are gpu1&gpu2 it will look like cs://gpu2.elgato.hpc.arizona.edu:11111

At this point we are now all set in order to start up the paraview gui in the remote desktop allocation. Again, please refer back to the Getting Started with Paraview documentation for instructions on this TODO replace this with the link in confluence.

Once the gui is open, configure a new connection, and put in just the server address ie gpu2.elgato.hpc.arizona.edu for the host, then supply the port as 11111

After about 10-30 seconds you will be connected and all of the filters and rendering will be accelerated by parallel multinode-mpi.

To double check whether your cores are being fully utilised consider using the following. First open a separate terminal ssh into each of the mpi nodes and run htop to see their individual cpu core utilization. As you interact with the paraview gui client the multinode-mpi server will undergo periods of low and high activity observable in the htop display.

Please consult the video below for any additional questions or reach out to the vislab-consult@list.arizona.edu

**Note that this video is using less mpi resources because waiting on two full elgato nodes wasn't allocating very quickly and made it harder to record the video. Just remember instead of -n 8 you can use -n 32 and then go out for a quick snack.**