Web access

Open OnDemand (OOD) web based interface is available for some clusters

while others have JupyterHub (JH)

Either of these lets you start a desktop session

Open OnDemand

Sign in with your alliance username and password. While the functionality is the same, menu layouts differ between clusters. The nibi menus are as follows

files: transfer files to/from cluster,
compute node: stuff that runs on compute nodes (i.e., in a job),
login node: stuff that runs on login nodes (i.e., immediately), and
my interactive sessions: connect to/cancel interactive compute stuff.

The desktop session is persistent. Closing the tab just closes the tab. It does not end the session. To reattach go to My Interactive Sessions and click Launch Desktop again. Pick Log out under the File menu in the menu bar along the top of the actual desktop to shut it down cleanly. Click Cancel on My Interactive Sessions to kill it.

Compute Node

Compute is stuff that happens on compute nodes. It runs in the context of a job and has whatever resources and time the job has been allocated. Here are general guidelines for specifying your allocation

account: ending in gpu if using a GPU (including hardware OpenGL) and cpu otherwise,
time: 8hrs or less to run in the interactive queue for faster start,
computers: one unless explicitly starting a multinode MPI process with srun or mpirun,
cores: enough cores so memory/core is 12G or less (regular memory queue) for faster start,
memory per computer: at least 8GB for a desktop session (see cores note about keeping memory/core 12G or less), and
gpu: if a GPU is required, the type the cluster has most of for faster start.

JupyterHub

Sign in with your alliance username and password. Enter your allocation details and click start. Here are general guidelines for specifying what your allocation

reservation: None unless taking part in a course (not this one) with a reservations,
account: ending in gpu if using a GPU (including hardware OpenGL) and cpu otherwise,
time: 3hrs or less to run in the interactive queue for faster start,
cores: enough cores so memory/core is 12G or less (regular memory queue) for faster start,
memory: at least 8GB for a desktop session (see cores note about keeping memory/core 12G or less),
oversubscription: select if possible (shares CPUs with other users) for faster start,
gpu configuration: if a GPU is required, the type the cluster has most of for faster start.
user interface: JupyterLab (Jupter Notebook is the older version before the renaming).

Once JupyterLab starts, click the Mate Desktop icon in the launcher window. This will open a desktop session in now browser tab.

The desktop session is persistent. Closing the tab just closes the tab. It does not end the session. Clicking the Desktop icon again will just reopen the existing one. To close it pick Log out in the menu bar along the top of the Desktop. The same is true of the entire JupyterLab session. Pick Log out under the File menu in the menu bar along the top to end it

Running Stuff

To get a terminal pick Mate terminal under Applications followed by System Tools in the menu bar along the top. The Compute Canada (the old name for the alliance) stack is loaded by default.

Clipboard

Browser limits remote access to clipboard for security reasons. With OOD, most modern browsers, appart from Firefox, will prompt for permission after which cutting and pasting between the desktop session and you desktop will just work.

For JupyterHub or OOD with Firefox, you have to first paste in or copy out of an intermediate cliboard. Under JupyterHub the intemediate clipboard is accessed at the top of the browser window (click on Remote Clipboard) and for OOD it is accessed on the right-hand side of the browser window (click on the pull out tab and then the clipboard icon).

Accelerated OpenGL

It is possible to use hardware accelerated OpenGL via VirtualGL if a whole GPU (i.e., not a multi-instance GPU) has been allocated. This can help applications showing poor interactivity when processing large scene graphs.

VirtualGL is enabled by default on OOD on nibi. For other clusters or JupyterHub it is necessary to start the program as follows

vglrun -d egl PROGRAM [ARGS ...]

JupyterLab Modules

JupyterLab is a python program, so it reloads some python modules that would not normally be loaded by default. If these conflict with other modules you need, just unload them or do a

module reset

before loading your desired modules.