The UAntwerp clusters have limited features for remote visualization on the login nodes of hopper and the visualization node of leibniz using a VNC-based remote display technology. On the regular login nodes of hopper, there is no acceleration of 3D graphics, but the visualisation node of leibniz is equipped with a NVIDIA M5000 card that when used properly will offer accelerated rendering of OpenGL applications. The setup is similar to the setup of the visualization nodes at the KU Leuven.

Using VNC turns out to be more complicated than one would think and things sometimes go wrong. It is a good solution for those who absolutely need a GUI tool or a visualization tool on the cluster rather than on your local desktop; it is not a good solution for those who don't want to invest in learning Linux properly and are only looking for the ease-of-use of a PC.

The idea behind the setup

2D and 3D graphics on Linux

Graphics (local and remote) on Linux-machines is based on the X Window System version 11, shortly X11. This technology is pretty old (1987) and nor really up to the task anymore with todays powerful computers yet has so many applications that support it that it is still the standard in practice (though there are efforts going on to replace it with Wayland on modern Linux systems).

X11 applications talk to a X server which draws the commands on your screen. These commands can go over a network so applications on a remote machine can draw on your local screen. Note also the somewhat confusing terminology: The server is the program that draws on the screen and thus runs on your local system (which for other applications will usually be called the client) while the application is called the client (and in this scenario runs on a computer which you will usually call the server). However, partly due to the way the X11 protocol works and partly also because modern applications are very graphics-heavy, the network has become a bottleneck and graphics-heavy applications (e.g., the Matlab GUI) will work sluggish on all but the fastest network connections.

X11 is a protocol for 2D-graphics only. however, it is extensible. Enter OpenGL, a standard cross-platform API for professional 3D-graphics. Even though its importance on Windows and macOS platforms had decreased as Microsoft and Apple both promote their own APIs (DirectX and Metal respectively), it is still very popular for professional applications and in the Linux world. It is supported by X11 servers through the GLX-extension (OpenGL for the X Window System). When set up properly, OpenGL commands can be passed to the X server and use any OpenGL graphics accelerator available on the computer running the X server. In principle, if you have a X server with GLX extension on your desktop, you should be able to run OpenGL programs on the cluster and use the graphics accelerator of your desktop to display the graphics. In practice however this works well when the application and X server run on the same machine, but the typical OpenGL command stream is to extensive to work well over a network connection and performance will be sluggish.

Optimizing remote graphics

The solution offered on the visualization node of leibniz (and in a reduced setting on the login nodes of hopper) consists of two elements to deal with the issues of network bandwidth and, more importantly, network latency.

VirtualGL is a technology that redirects OpenGL commands to a 3D graphics accelerator on the computer where the application is running or to a sofware rendering library. It then pushes the rendered image to the X server. Instead of a stream of thousands or millions of OpenGL commands, one large image is now passed over the network to the X server, reducing the effect of latency. These images can be large though, but with an additional piece of software on your client, called the VGL client, VirtualGL can send the images in compressed form which strongly reduces the bandwidth requirements. To use VirtualGL, you have to start the OpenGL application through the vglrun-command. That command will set up the application to redirect OpenGL calls to the VirtualGL libraries.

VirtualGL does not solve the issue of slow 2D-rendering because of network latency and also requires the user to set up a VGL client and an X server on the local desktop, which is cumbersome for less experienced users. We solve this problem through VNC (Virtual Network Computing). VNC consists of three components: a server on the computer where your application runs, a client on your desktop, and a standardized protocol for the communication between server and client. The server renders the graphics on the computer on which it runs and sends compressed images to the client. The client of course takes care of keyboard and mouse input and sends this to the server. A VNC server for X applications will in fact emulate a X server. Since the protocol between client and server is pretty standard, most clients will work with most servers, though some combinations of client and server will be more efficient because they may support a more efficient compression technology. Our choice of server is TurboVNC which is maintained by the same group that also develops VirtualGL and has an advanced implementation of a compression algorithm very well suited for 3D graphics. TurboVNC has clients for Windows, macOS and Linux. However, our experience is that it also works with several other VNC clients (e.g., Apple Remote Desktop), though it may be a bit less efficient as it may not be able to use the best compression strategies.

The concept of a Window Manager

When working with Windows or macOS, we're used to seeing a title bar for most windows with buttons to maximize or hide the window, and borders that allow to resize a window. You'd think this functionality is provided by the X server, but in true UNIX-spirit of having separate components for every bit of functionality, this is not the case. On X11, this functionality is provided by the Window Manager, a separate software package that you start after starting the X server (or may be started for you automatically by the startup script that is run when starting the X server). The basic window managers from the early days of X11 have evolved into feature-rich desktop enviroments that do not only offer a window manager, but also a task bar etc. Gnome and KDE are currently the most popular desktop environments (or Unity on Ubuntu, but future editions of Ubuntu will return to Gnome). However, these require a lot of resources and are difficult to install on top of TurboVNC. Examples of very basic old-style window managers are the Tab Window Manager (command twm) and the Motif Window Manager (command mwm). (Both are currently available on the login nodes of hopper.)

For the remote visualization setup on the UAntwerp clusters, we have chosen to use the Xfce Desktop Environment which is definitely more user-friendly than the rather primitive Tab Window Manager and Motif Window Manager, yet requires less system resources and is easier to set up than the more advanced Gnome and KDE desktops.

Prerequisites

You'll need a ssh client on your desktop that provides port forwarding functionality on your desktop. We refer to the "Access and data transfer" section of the documentation on the user portal for information about ssh clients for various client operating systems. PuTTY (Windows) and OpenSSH (macOS, Linux, unix-compatibility environment on Windows) both provide all required functionality.

Furthermore, you'll need a VNC client, preferably the TurboVNC client.

Windows

We have tested the setup with three different clients:

  • The TurboVNC client can be downloaded by following the Download link on the TurboVNC web site (which at the moment of writing this documentation takes you to a sourceforge page). Binaries are available for both 32-bit and 64-bit windows systems. This client is made by the same people as the server we use so in theory one should expect the least problems with this setup.
  • TigerVNC is a client whose development is supported by the Swedish company Cendio who makes a remote display server product (ThinLinc) based on TigerVNC. Binaries for 32-bit and 64-bit windows (vncviewr-*.*.*.exe) can be downloaded by following the link on the GitHub Releases page. These binaries are ready-to-run.
  • ThightVNC is also a popular free VNC implementation. 32-bit and 64-bit Windows installers can be downloaded from the download page on their website. When installing on your PC or laptop, make sure to chose the "custom install" and only install the TightVNC Viewer.

All three viewers are quite fast and offer good performance, even when run from home over a typical broadband internet connection. TigerVNC seems to be a bit quicker than the other two, while TightVNC doesn't allow you to resize your window. With the other two implementations, when you resize your desktop window, the desktop is also properly resized.

macOS

Here also there are several possible setups:

  • The TurboVNC client can be downloaded from the TurboVNC web site. The macOS client is Java-based. Packages are available for both Apple Java on older versions of OS X and Oracle Java (which you will need to install if it is not yet on your system). We advise to use the Oracle Java version as Java needs frequent security updates and Apple Java is no longer maintained.
  • TigerVNC, a client whose development is supported by the Swedish company Cendio who makes a remote display server product (ThinLinc) based on TigerVNC, is a native macOS client. At the time of writing (version 1.9.0), it is still only distributed as a 32-bit binary so you may get warnings on some versions of macOS. However, there already exist 64-bit pre-release builds so future versions will certainly fully support future macOS versions. Some places report that this client is a lot slower than the the TurboVNC one on macOS.
    Binaries are available. Look for the tigervnc-*.dmg files, which contrary to those for Windows and Linux, only contain the viewer software.
  • A not-so-good alternative is to use the Apple Screen Sharing feature which is available through the Finder (command-K key combination) or Safari (URL bar) by specifying the server as a URL starting with svn://. This VNC client is considerably slower though than the TurboVNC client, partly because it doesn't support some of the TurboVNC-specific compression algorithms.

Linux

RPM and Debian packages for TurboVNC can be downloaded from the TurboVNC web site and are available in some Linux distributions. You can also try another VNC client provided by your Linux distribution at your own risk as we cannot guarantee that all VNC viewers (even recent ones) work eficiently with TurboVNC.

How do I run an application with TurboVNC?

Running an application with TurboVNC requires 3 steps:

  • Start the VNC server on the cluster
  • Start the VNC client on your desktop/laptop and connect to the server
  • Start your application

Starting the server

  1. Log on in the regular way to one of the login nodes of hopper or to the visualization node of Leibniz. Note that the latter should only be used for running demanding visualizations that benefit from the 3D acceleration. The node is not meant for those who just want to run some lightweight 2D Gui application, e.g., an editor with GUI.
  2. Load the module vsc-vnc:
    module load vsc-vnc
    This module does not only put the TurboVNC server in the path, but also provides wrapper scripts to start the VNC server with a supported window manager / dekstop environment. Try module help vsc-vnc for more info about the specific wrappers.
  3. Use your wrapper of choice to start the VNC server. We encourage to use the one for the Xfce desktop environment:
    vnc-xfce
  4. The first time you use VNC, it will ask you to create a password. For security reasons, please use a password that you don't use for anything else. If you have forgotten your password, it can easily be changed with the vncpasswd command and is stored in the file ~/.vnc/passwd in encrypted form. It will also ask you for a viewer-only password. If you don't know what this is, you don't need it.
  5. Among other information, the VNC server will show a line similar to:
    Desktop 'TurboVNC: viz1.leibniz:2 (vsc20XXX)' started on display viz1.leibniz:2
    Note the number after TurboVNC:viz1.leibniz, in this case 2. This is the number of your VNC server, and it will in general be the same as the X display number which is the last number on the line. You'll need that number to connect to the VNC server.
  6. It is in fact safe though not mandatory to log out now from your SSH session as the VNC server will continue running in the background.

The standard way of starting a VNC server as described in the TurboVNC documentation is by using the vncserver command. However, you should only use this command if you fully understand how it works and what it does. Also, please don't forget to kill the VNC server when you have finished using it as it will not be killed automatically when started through this command (or use the -autokill command line option at startup). The default startup script (xstartup.turbovnc) which will be put in the ~/.vnc directory on first use does not function properly on our systems. We know this and we have no intent to repair this as we prefer to install the vncserver command unmodified from the distribution and provide wrapper scripts instead that use working startup files.

Connecting to the server

  1. In most cases, you'll not be able to connect directly to the TurboVNC server (which runs on port 5900 + the server number, 5902 in the above example) but you will need to create a SSH tunnel to forward traffic to the VNC server. The exact procedure is explained in length in the pages "Creating a SSH tunnel using PuTTY" (for Windows) and "Creating a SSH tunnel using OpenSSH" (for or Linux and macOS) .
    You'll need to tunnel port number (5900 + server number) (5902 in the example above) on you local machine to the same port number on the node on which the VNC server is running. You cannot use the generic login names (such as login.hpc.uantwerpen.be) for that as you may be assigned a different login node as you were assigned just minutes ago. Instead, use the full names for the specific nodes, e.g., login1-hopper.uantwerpen.be, login2-leibniz.uantwerpen.be or viz1-leibniz.uantwerpen.be.
    1. In brief:With OpenSSH, your command will look like
      ssh -L 5902:viz1-leibniz.uantwerpen.be:5902 -N vsc20XXX@viz1-leibniz.uantwerpen.be
    2. In PuTTY, select "Connections - SSH - Tunnel" in the left pane. As "Source port", use 5900 + the server number (5902 in our example) and as destination the full name of the node on which the VNC server is running, e.g., viz1-leibniz.uantwerpen.be.
  2. Once your tunnel is up-and-running, start your VNC client. The procedure depends on the precise client you are using. However in general, the client will ask for the VNC server. That server is localhost:x where x is the number of your VNC server, 2 in the above example. It will then ask you for the password that you have assigned when you first started VNC.
  3. If all went well, you will now get a window with the desktop environment that you have chosen when starting the VNC server
  4. Do not forget to close your tunnel when you log out from the VNC server. Otherwise the next user might not be able to connect.

Note that the first time that you start a Xfce session with TurboVNC, you'll see a panel "Welcome to the first start of the panel". Please select "Use default config" as otherwise you get a very empty desktop.

Starting an application

  1. Open a terminal window (if one was not already created when you started your session).
    In the default Xfce-environment, you can open a terminal by selecting "Terminal Emulator" in the "Applications" menu in the top left. The first time it will let you chose between selected terminal applications.
  2. Load the modules that are required to start your application of choice.
  3. 2D applications or applications that use a sofware renderer for 3D start as usual. However, to start an application using the hardware-accelerated OpenGL, you'll need to start it through vglrun. Usually adding vglrun at the start of the command line is sufficient.
    This however doesn't work with all applications. Some applications require a special setup.
    1. Matlab: start matlab with the -nosoftwareopengl option to enable accelerated OpenGL:
      vglrun matlab -nosoftwareopengl
      The Matlab command opengl info will then show that you are indeed using the GPU.
  4. When you've finished, don't forget to log out (when you use one of our wrapper scripts) or kill the VNC server otherwise (using vncserver -kill :x with x the number of the server).

Note: For a quick test of your setup, enter

vglrun glxinfo
vglrun glxgears

The first command will print some information about the OpenGL functionality that is supported. The second command will display a set of rotating gears. Don't be fooled if they appear to stand still but look at the "frames per second" printed in the terminal window.

Common problems

  • Authentication fails when connecting to the server: This happens occasionaly when switching between different versions of TurboVNC. The easiest solution is to simply kill the VNC server using vncserver -kill :x (with x the display number), set a new VNC password using vncpasswd and start over again.
  • Xfce doesn't show the task bar at the top of the screen: This too happens sometimes when switching between versions of Xfce4, or you may have screwed up your configuration in another way. Remove the .config/xfce-centos7 directory (rm -r .config/xfce-centos7) or the .config/xfce-sl6 directory depending on whether you are working on a CentOS7 system (Leibniz curently) or Scientific Linux 6 system (/hopper currently), kill the VNC server and start again.

Links

Components used in the UAntwerp setup

Related technologies

Client

Systems