Finding fast, user-friendly ways to organize and analyze all this data is the job of the OLCF Advanced Data and Workflow Group and computer scientists like Benjamín Hernández, who has developed a new visualization tool called SIGHT for OLCF users.
"The amount of data users deal with is huge, and we want them to be able to easily visualize their datasets remotely and in real-time to see what they are simulating", Benjamín Hernández stated. "We also want to provide 'cinematic rendering' to enhance the visual perception of visualizations."
Through scientific visualizations, researchers can better compare experimental and computational data. Using a type of scientific visualization known as exploratory visualization, researchers can interactively manipulate 3D renderings of their data to make new connections between atomic structure and physical properties, thereby improving the effectiveness of the visualization.
However, as scientific data grow in complexity, so too do memory requirements - especially for exploratory visualization. To provide an easy-to-use, remote visualization tool for users, Benjamín Hernández developed SIGHT, an exploratory visualization tool customized for OLCF user projects and directly deployed on OLCF systems.
As opposed to traditional visualization in which images are often rendered during post-processing, exploratory visualization can enable researchers to improve models before starting a simulation; make previously unseen connections in data that can inform modelling and simulation; and more accurately interpret computational results based on experimental data.
Benjamín Hernández incrementally developed the exploratory visualization tool SIGHT by working with a few teams of OLCF users to fold in the specific features they needed for their projects.
To study how lasers transform metal surfaces to create complex, multiscale roughness and drive the ejection of nanoparticles, a team led by materials scientist Leonid Zhigilei of the University of Virginia used Titan to simulate more than 2 billion atoms over thousands of time steps.
"The initial attempts to visualize the atomic configurations were very time-consuming and involved cutting the system into several pieces and reassembling the images produced for different pieces", Leonid Zhigilei stated. "SIGHT, however, enabled the researchers at the University of Virginia to take a quick look at the whole system, monitor the evolution of the system over time, and identify the most interesting regions of the system that require additional detailed analysis."
SIGHT provides high-fidelity "cinematic" rendering that adds visual effects - such as shadowing, ambient occlusion, and photorealistic illumination - that can reveal hidden structures within an atomistic dataset. SIGHT also includes a variety of tools, such as a sectional view that enables researchers to see inside the chunks of atoms forming the dataset. From there, the research team can further explore sections of interest and use SIGHT's remote capabilities to present results on different screens, from mobile devices to OLCF's Powerwall, EVEREST.
SIGHT also reduces the amount of work users must wade through upfront. Out-of-the-box tools come with many customization options and an underlying data structure that adapts the product to run on a variety of systems but slows performance by taking up more memory.
Benjamín Hernández has so far deployed SIGHT on Rhea, OLCF's 512-node Linux cluster for data processing, and DGX-1, a NVIDIA artificial intelligence supercomputer with eight GPUs. On both Rhea and DGX-1, SIGHT takes advantage of high-memory nodes and optimizes CPU and GPU rendering through the CPU-rendering OSPRay and GPU-rendering NVIDIA OptiX libraries.
Keeping data concentrated on one or a few nodes is critical for exploratory visualization, which is possible on machines like Rhea and DGX-1.
"With traditional visualization tools, you might be able to perform exploratory visualization to some extent on a single node at low interactive rates, but as the amount of data increases, visualization tools must use more nodes and the interactive rates go down even further", Benjamín Hernández stated.
To model fundamental energy processes in the cell, another user project team led by chemist Abhishek Singharoy of Arizona State University simulated 100 million atoms of a membrane that aids in the production of adenosine triphosphate, a molecule that stores and transports energy in the cell. Using Rhea, collaborator Noah Trebesch from Emad Tajkhorshid's laboratory at the University of Illinois at Urbana-Champaign then used SIGHT to extended molecular visualization of biological systems to billions of atoms with a model of a piece of the endoplasmic reticulum called the Terasaki ramp.
Compared to a typical desktop computer with a commodity GPU that a researcher might use for running SIGHT at their home university or institution, using SIGHT for remote visualization on Rhea enabled higher particle counts and frame rates by several orders of magnitude. With 1 terabyte of memory available in a single Rhea node, SIGHT using the OSPRay backend to reveal over 4 billion particles and there is still memory available for larger counts.
Furthermore, running SIGHT on the DGX-1 system with the NVIDIA OptiX backend resulted in frame rates up to 10 times faster than a typical desktop computer and almost 5 times faster than a Rhea node.
Anticipating the arrival of Summit later this year, Benjamín Hernández is conducting tests on how remote interactive visualization workloads can be deployed on the OLCF's next-generation supercomputer.