Muhammad Usman
Research Engineer
BSC
Abstract:
Data Processing Units (DPU) are a new class of programmable processor system on a chip (SoC) that combine industry-standard, high-performance, software-programmable, and processing elements (such as CPUs and GPUs) tightly coupled to the other SoC components.
DPUs are oriented to infrastructure computing platforms for processing software-defined networking, storage, and cybersecurity operations. DPUs can combine powerful computing, high-speed networking, and extensive programmability to deliver hardware-accelerated solutions for the most demanding workloads.
The main manufacturer of DPUs is NVIDIA, which provides them with a specific software development kit (SDK) for their BlueField DPU devices. DOCA, the SDK, is designed to unlock data center innovation by enabling the rapid creation and deployment of applications and services. DOCA provides industry-standard open application programming interfaces (API) and frameworks for, among others, networking, security, storage, high-performance computing (HPC), or artificial intelligence (AI) applications. The frameworks simplify application offload with integrated NVIDIA acceleration packages.
Despite the provided APIs, DOCA still presents an entry barrier for new users, especially HPC programmers, who are not likely to be familiar with the novel programming paradigm. OpenMP DPU Offloading Support (ODOS) framework is presented to overcome this barrier by providing DOCA-based offloading through the standard OpenMP syntax.
Unfortunately, DPU devices are not widely available in production supercomputing clusters today. Therefore, we implemented pDOCA, an emulation layer for programming using DOCA libraries without access to the DPU hardware. pDOCA uses the LD_PRELOAD mechanism of Linux to trap DOCA calls. These calls can then be translated to any other backend, such as TCP or UCX, for actual implementation. pDOCA has been specifically implemented to enable ODOS emulation for users without DPU devices. The infrastructure is composable and can be extended to improve more projects in the future.
More information: hpckp.org/
22 авг 2024