Developing and deploying applications across heterogeneous infrastructures like HPC or Cloud with diverse hardware is a complex problem. Enabling developers to describe the application deployment and optimising runtime performance while protecting data privacy and security is paramount. SODALITE, a Horizon 2020 project, aims to solve this by providing tools for increasing design and runtime effectiveness of software-defined infrastructures.

SODALITE targets complex applications and workflows that are deployed on heterogeneous environments such as virtual machines, containerized HPC clusters, Cloud and Edge devices. In this context, deploying the application in infrastructures usually requires command-line-based access and expert knowledge in the application domain. Even further, optimising HPC applications also requires a good background in networking technologies and parallel programming. With SODALITE, there is a stack on top of the low-level layer which makes it easier for non-experts to use VMs, HPC (with diverse hardware) and get optimal performance.

SODALITE allows the Application Ops Expert (AoE) to model the deployment of an optimised application on an infrastructure target using application, infrastructure and performance abstractions defined by resource experts. The models created by the AoE are automatically translated into infrastructural code, which is then translated to an optimised deployment using state of the art container technologies. This optimized application in a container is then deployed by an orchestrator on multiple diverse computing platforms. Deployment to clusters and supercomputers with homogeneous or heterogeneous node architectures for heavy batch computations, including resources available on the Cloud and Edge devices, is supported.

In optimising the application deployment, SODALITE attempts to bring the vast knowledge of performance optimisation acquired by the HPC industry over decades into the cloud computing area. The SODALITE deployment framework enables automated performance optimization before deployment (static) as well as at runtime (dynamic). Application performance optimization is very dependent on the application, its configuration and the infrastructure. SODALITE focuses on supporting three major application types for static optimisation: AI training/Inference, Big Data Analytics and Traditional HPC. Traditional HPC refers to applications using standards like MPI and OpenMP, AI training /Inference refers to applications that use computational graphs to extract features from data and Big data refers to  data processing applications at a larger scale.

SODALITE at ISC-HPC

As part of the now virtual ISC-HPC proceedings, SODALITE is presenting a poster. SODALITE had planned a booth at the physical fair, but is now preparing an innovative remote booth experience, and will transition to digital most of the elements that would be included at a traditional booth, including the insightful interactions, live demonstrations and joyful competitions. More information forthcoming at sodalite.eu/events.