RENO, Nev. — August 28, 2024 — Software infrastructure company CIQ today released the Fuzzball computing and data management platform for performance-intensive computing (PIC). CIQ said it is shipping new capabilities in Fuzzball, giving individual researchers tools designed to access PIC resources faster and without having to learn and manage infrastructure provisioning and management.
CIQ said Fuzzball is research ready, with pre-built workflow templates that allow researchers to work with the tools they already depend upon. Examples include PyTorch, Stable Diffusion and Tensorflow for AI; OpenRadioss and GROMACS for simulation and modeling; RStudio Server and Jupyter Notebooks for programming; and MATLAB for broad scientific use cases.
Legacy high performance computing infrastructure doesn’t stand up to today’s demands for PIC, CIQ said. The expansion of artificial intelligence and PIC into the enterprise has exacerbated the issue. End users (engineers, scientists, researchers, analysts, etc.) are running diverse and increasingly large and complex workloads that demand exponentially more computing power; modern, specialized hardware; and more optimized user interfaces. Today’s researchers are also running applications that benefit from containerization and rely on application interfaces that legacy systems were not designed to accommodate. Modern researchers also need PIC that performs well and securely across geographical distances, with easy portability and reproducibility.
In short, legacy HPC infrastructure does meet modern technical and business expectations and requirements for hardware, software and geographical workload distribution. Also, it lacks many of the critical advantages of cloud computing and automation and requires computer science expertise to use and manage. It can even require specialized knowledge to perform minimal functions like logging in. In contrast, Fuzzball is a modern, PIC platform that helps researchers maximize speed to innovation, no computer science degree required.
With Fuzzball, scientists and innovators no longer need specialized knowledge of computer systems to run their workloads. Fuzzball’s graphical user interface (GUI) abstracts away the complexity of workflow design and execution, helping researchers and engineers in a variety of fields and industries to rapidly deploy a workload without computer science expertise.
Fuzzball also eliminates the time innovators spend defining their research workloads, offering example workflows to help users get started. In addition, Fuzzball’s standardized, human-readable workload file format enables users to copy and paste workload files from one Fuzzball cluster to another without any rework.
“Over the two past years, CIQ has been working with customers in a limited preview of Fuzzball across enterprise and public sectors,” said Gregory Kurtzer, CEO of CIQ. “We heard from customers and the community that researchers are burdened with becoming Linux and infrastructure experts in order to do their research and this is affecting their productivity and ability to meet their demands of innovation. While organizations cumulatively have spent many billions of dollars for their computing resources, these systems are based on a legacy architecture and no longer meet the demands of today’s advanced use-cases. HPC architecture has been the same for the past 30 years; that all changes with Fuzzball.”
Traditional container orchestration platforms like Kubernetes assume that containerized processes are expected to run forever. This is great when building cloud-native web apps with microservice architecture, but the design falls over when you attempt to orchestrate jobs that have logical conclusions and depend on one another in unique ways.
In contrast, Fuzzball allows customers to build computing-focused workflows consisting of jobs pipelined together, all operating within their respective containerized environments. Job pipelines are constructed as acyclic graphs, and as a result pipelines can be linear, parallel or any combination of the two. Jobs themselves can also be serial, multi-system parallel (MPI, GasNet, etc.) or arrays. Customers describe the containers, resources and requirements that each job needs as well as the relationship between jobs. Jobs are submitted and monitored by the click of a button.
Another unique aspect of Fuzzball’s architecture is its approach to data management. Rather than depend on manual data movement, lifecycle and staging, Fuzzball automates data management while maintaining security policies and auditability. Fuzzball will ingress the required data for a given workflow into a file-based volume, and after the job completes, the resulting identified data can be egressed back into external file or object-based storage.
Fuzzball abstracts the infrastructure and automates the process of running complex workflows which drives efficiency of innovation, development and individuals while removing infrastructure-specific dependencies enabling greater levels of portability, cost effectiveness and collaboration.
