What a workstation is
A workstation is a server in a cloud datacenter that streams a desktop to you. That one property explains most of how workstations behave, including what persists and what does not.
The model
Rescale provisions a virtual machine in a cloud region you choose, installs the software you asked for, and starts a remote-display server on it. Your browser or a local client receives a video stream of that desktop and sends back your keyboard and mouse input.
Two consequences follow, and they account for most surprises:
- Latency is a network property, not a hardware one. A faster workstation doesn't feel more responsive if the datacenter is far away. Region choice matters more than core count for interactive feel.
- The machine's disk is not your disk. Files live on the workstation until something deliberately copies them back to Rescale. See saving and syncing files — this is the most common way people lose work.
Workstation or job?
A workstation is not a job. Jobs have one purpose, to solve. Workstations are for the three things you can't do in a batch job:
- Pre-processing, building geometry, meshing, setting up the model.
- Post-processing, visualizing results. In practice this is the largest use, roughly 70% of workstation time.
- Interactive solving, driving a solver through its GUI, pausing it, changing a parameter, and continuing. This consumes the most core-hours per session.
If your model is already set up and you only need results, submit a job. A workstation costs money while it sits open; a job stops charging when it finishes.
What you pay for
Three components, all billed per core-hour:
The workstation license covers the visualization capability, so it's charged only on the node actually running the desktop. On a multi-node workstation, worker nodes carry just the platform license.
Suspending stops the compute charge, but the platform license keeps accruing on every node. That's what pays for maintaining your saved disk. Suspending is much cheaper than running, not free. To stop all charges, stop the workstation. See suspend and resume.
Single-node and multi-node
Most workstations are a single machine. A multi-node workstation is a cluster: a head node running the desktop, plus worker nodes behind it, so you can drive a parallel solve from the GUI.
The tradeoff is that multi-node workstations cannot be suspended, suspension is single-node only. On a multi-node session, closing down means stopping, and stopping means the disk goes away.
Latency and what improves it
Expect under 100 ms in-browser. Two changes get you under 50 ms:
- Use the native client instead of the browser. The client connects straight to the workstation; the browser routes through the Rescale platform first. If you're in India using the EU platform to run a workstation in India, the browser path goes to Germany and back. The client doesn't.
- Pick a nearby region.
A GPU coretype also matters for anything graphics-heavy: a GPU intended for visualization renders the frames the stream is made of. See connecting for the client setup and limits for what's supported.
Start-up, and why it's slow
A first launch takes about 15 minutes for Linux and 25 for Windows. That time isn't idle: Rescale provisions the cluster, installs the OS, applies security updates, attaches a volume containing your software, and runs a configuration script for it.
Windows is slower for a structural reason. Ideally software is self-contained on its volume and just needs mounting. But much Windows software installs via an installer that writes to the registry and registers background services, changes that can't be replayed by attaching a disk. So the configuration step effectively runs the installer from scratch.
This is what makes suspension valuable. Resuming a suspended workstation takes under 3 minutes, because the software is already installed on the saved disk. If you'll use a workstation again, suspend it rather than stopping it.
Related
- Create a workstation — the setup flow
- Suspend and resume — avoiding the 15–25 minute start
- Save and sync your files — what survives shutdown
- Limits & supported configurations — OS, clouds, protocols, hardware
- Jobs — the batch alternative