UPF Scientific Computing Core Facility
Scientific viewers and interactive 3D tools on Correfoc, where parallel storage, memory and GPU resources work together. Keep the data on the cluster and bring the interactive view to your own high-resolution screen.
Start with the data
Large-data visualization depends on feeding the application as well as drawing the image. For microscopy, the storage and memory path can matter just as much as the GPU.
A microscopy time-lapse can combine three spatial dimensions (X, Y and Z) with time (T). Depending on resolution and bit depth, a single image plane can occupy tens or hundreds of megabytes. Navigating Fiji’s Z and T sliders may request a succession of these planes; a projection or 3D view may need an entire volume at each time point.
Correfoc’s GPFS parallel filesystem can distribute file data across storage resources and feed it to the computing nodes over the HPC data network. Sustained reads help keep playback and Z/T navigation responsive. If loading falls behind, the viewer can pause while waiting for the next image.
The parallel filesystem combines shared project capacity and aggregate I/O bandwidth with server memory and CPU/GPU processing. Work directly on large research datasets, without staging another copy on each workstation or making its local disk the only path to the data.
The browser receives compressed display video, rather than every raw microscopy plane. Storage traffic inside Correfoc and the network bitrate of your remote view are separate quantities. Your existing permissions apply to files in homes, projects and scratch.
From the data to your display
Explore images and read fine interfaces on a high-definition screen. Choose compatible display and video settings for the task, from interactive motion to close inspection.
Spatial detail
Use a detailed remote workspace at 4K (3840 × 2160), or 5K (5120 × 2880) where the session, codec, browser and display support it. HiDPI scaling keeps interface elements usable while making more screen pixels available for the scientific view.
Tonal precision
Compatible HEVC and AV1 profiles on the L40S can encode video at 10 bits per component: up to 1024 code values rather than 256 at 8 bits. The extra coding precision can help reduce banding in gradients, when supported by the receiving browser.
Fine colour detail
Compatible H.264 or HEVC profiles offer 4:4:4 chroma for coloured text, fine lines and scientific interfaces. AV1 on the L40S uses 4:2:0. Resolution, bit depth and chroma sampling describe different properties of the streamed image.
The video format is separate from the depth of the original dataset. A 16-bit image remains 16-bit data in the application; streaming its desktop does not transmit all those values as raw pixels. Current capture paths can start from an 8-bit desktop, so 10-bit encoding adds coding precision without recovering source levels that were never captured. It also does not, by itself, provide HDR. Available 10-bit formats depend on the codec and transport that the browser actually decodes.
Motion and inspection
Scientific work alternates between movement and close inspection. A remote session should handle both, from a rotating 3D model to the labels and fine lines you stop to read.
01 · During motion
Rotate a model, scroll through a microscopy stack or navigate a simulation. The stream adapts its bitrate to the network to keep interaction responsive.
02 · At rest
When the view settles, compatible streaming profiles progressively refine the held image, including under controlled bitrate. A blinking caret or another small changing area need not prevent the rest of the image from becoming clearer.
03 · Back to motion
In updated streaming profiles, bitrate recovery stays tied to measured link capacity during quiet periods, helping limit a burst of queued video when you next rotate, scroll or zoom.
Still-image refinement is paced by the available link. It improves the streamed view; it does not change the source dataset.
Display damage or content hashes identify what changed; compatible NVENC paths check content on the GPU. Change-aware modes can suppress redundant updates, while continuous-video modes may keep encoding frames. A held view can still be refined in either mode.
Choose how you interact
High-resolution video brings the view to you. Keyboard, pointer and touch input make the session interactive in the other direction.
Touchscreen gamepad
An on-screen gamepad gives phone and tablet users virtual sticks and action buttons over the remote view. Compatible applications receive game-controller input without requiring a physical controller.
Use touch controls to navigate prepared 3D scenes or interact with simulations. Mouse, keyboard, touch and controller commands return to the application on Correfoc; camera video is a separate, optional uplink.
Extend a compatible remote desktop to a second browser window. Keep code on one monitor and images, plots or 3D results on the other.
Prepared workflows can receive camera and microphone input, play application audio and use a physical joystick or gamepad. Combine live visual feedback with computer vision or simulation.
Where sharing is enabled, authorised collaborators can join the same display with the permitted viewing or control role. Clipboard and selected file transfers connect the session with your local work.
Research workflows
Examples range from everyday analysis to environments prepared for advanced imaging, simulation and robotics. Contact the SCC team to check software availability, licensing and the resources your workflow needs.
Develop and debug code, explore datasets and monitor analyses before moving longer runs into scheduled jobs.
Jupyter · RStudio · VSCode · MATLAB · Stata · TensorBoard
Open image stacks, inspect segmentations and annotate microscopy data using the memory and storage available on Correfoc.
Fiji · napari · QuPath · CellProfiler · Cellpose · ilastik
Inspect volumetric images, segment structures and explore scientific imaging results through graphical interfaces.
3D Slicer · MITK · ITK-SNAP
Explore molecular structures and trajectories, inspect biological networks and browse genomic data close to project storage.
ChimeraX · PyMOL · VMD · Cytoscape · IGV
Prepared environments
Work with 3D geometry and navigate interactive scenes rendered remotely, using browser input and supported controllers.
Blender · Unreal Engine
Prepared environments
Connect camera input to a vision experiment or combine robot simulation, visual feedback and joystick commands in a configured workflow.
Camera-based inference · ROS 2 · Isaac Sim
The technology behind the experience
GPU capture and hardware video encoding support the experience described above. They work alongside the parallel filesystem, application memory, rendering and the client.
Compatible zero-copy paths pass a GPU image buffer to the encoder without a pixel readback through CPU memory. These include DMA-BUF on Wayland and DRI3 or NvFBC on X11, when rendering, capture and encoding can share the buffers.
Other configurations use a readback path. Keeping this stage short helps the streaming pipeline, while application rendering and network conditions still contribute to the overall latency.
Each L40S has three NVENC engines. Compatible AV1 profiles split one captured screen frame into three horizontal bands, encode those bands in parallel, and produce a single video stream. Dedicated hardware handles video encoding while the application uses the allocated computing resources.
One captured screen frame
Connectivity
The access portal starts the allocated session. Its graphics can then reach the browser through a real-time media route or through the web connection.
Media follows an ICE-selected path, normally using UDP where reachable, with relay or TCP options where configured. UDP avoids TCP head-of-line blocking during packet loss. The media route is separate from the portal’s HTTP proxy; signaling still uses the web connection.
Encrypted WebSockets carry video, audio and interaction over TCP through a compatible web proxy. Video is decoded through WebCodecs; a JPEG display path can serve browsers without a suitable video decoder. It fits networks where a separate media path is unavailable.
Transport and codec are separate choices. AV1 can be delivered over either transport when the browser supports it. Network policy and the enabled session profile determine which options are offered.
Getting started
Follow the SCC access requirements, then choose an environment and a resource allocation for your research. Interactive sessions have an allocated CPU, memory, optional GPU resources and a time limit.
Select a notebook, application or desktop suited to your workflow.
Choose the available CPU, memory, GPU and session-duration options.
Open the allocated session from your browser and work on your data.
Save relevant results in project storage and end the session when finished.
Discuss your application, dataset, display and network requirements with the SCC team.
