BSC Democratizes the Use of NVIDIA BlueField DPUs in High-Performance Computing with ODOS: OpenMP offloading support for NVIDIA DOCA

16 November 2023

The Accelcom research group, led by researcher Antonio Peña, has developed a solution that provides support for standard OpenMP offloading syntax and semantics for DPUs

Data processing units (DPUs) have emerged as a groundbreaking class of programmable processor system on a chip (SoC), delivering accelerated solutions for the most demanding workloads. DPUs combine industry-standard, high-performance, software-programmable processing elements such as CPUs and GPUs, tightly integrated with robust networking and other SoC components.

These DPUs are specifically designed for infrastructure computing platforms, empowering software-defined networking, storage, and cybersecurity operations. For such applications, NVIDIA offers the NVIDIA BlueField DPU platform along with a comprehensive software framework called DOCA™. NVIDIA DOCA unlocks data center innovation by facilitating the rapid creation and deployment of applications and services on the BlueField DPU. It provides industry-standard open application programming interfaces (APIs) and services for networking, security, storage, high-performance computing (HPC), and artificial intelligence (AI) applications, streamlining application offload with integrated NVIDIA acceleration packages.

To further enhance accessibility to DPUs, particularly for HPC programmers unfamiliar with this novel programming paradigm, Antonio J. Peña, leader of the AccelCom research group at the Barcelona Supercomputing Center (BSC), spearheaded an initiative. Recognizing the opportunity to democratize the use of DPUs in HPC, Peña established a working group comprising Muhammad Usman (Research Engineer) and Sergio Iserte (Senior Researcher). Their mission was to develop a solution that provides support for standard OpenMP offloading syntax and semantics for DPUs, aptly named ODOS (ODOS (OpenMP offloading support for NVIDIA DOCA)).

ODOS extends the LLVM compiler infrastructure to enable OpenMP offloading for NVIDIA BlueField DPUs. With ODOS, users can seamlessly offload computations to GPUs and/or DPUs by using the familiar and widely adopted standard OpenMP syntax. The solution eliminates barriers for new users, enabling them to leverage the power of DPUs within their existing programming frameworks.

The impact of ODOS has not gone unnoticed. NVIDIA has provided substantial assistance and resources for ODOS, funding its development and facilitating its promotion among its user community.

Antonio J. Peña said: “The AccelCom group members at BSC have a long tradition of easing the of use of HPC accelerators and network communications to domain developers. We are excited to see the interest of the community in ODOS, the first of its kind. I expect to soon see many more applications benefiting from the performance advantage of NVIDIA BlueField DPUs thanks to the seamless interface provided by ODOS. We are already collaborating with NVIDIA on new functionalities and enhanced performance.”

BSC’s development of ODOS for NVIDIA BlueField DPUs aims to boost HPC and scientific computing applications that can help tackle various challenges,” said Gilad Shainer, senior vice president of networking at NVIDIA. “By integrating DPUs seamlessly into the OpenMP ecosystem, we are empowering programmers with greater access to accelerated computing, unlocking new possibilities for high-performance applications.”

With the collaborative efforts of BSC, NVIDIA, and the ODOS project team, the HPC community can now harness the full potential of DPUs and leverage their capabilities within the established OpenMP framework. ODOS has been recently open-sourced and will be presented during the SC23 conference as part of a workshop paper and a tutorial. Planned upcoming steps include upstream contribution to the LLVM project and MPI integration (i.e., supporting MPI calls within offloaded tasks).


Caption: Antonio J. Peña and Sergio Iserte, part of the ODOS team, presented ODOS at SC23