An open-source framework for offloading big data and AI tasks (Offload) to heterogeneous compute units
No Thumbnail Available
Meeting name
Sponsors
Date
Journal Title
Format
Thesis
Subject
Abstract
The ever-increasing demands of artificial intelligence (AI) and big data processing have spurred the rapid development of novel hardware architectures specifically designed for computationally intensive tasks. Alongside these advancements, software solutions are emerging to exploit this specialized hardware by offloading tasks. However, proprietary software often necessitates a substantial learning curve for users, hindering widespread adoption and flexibility. This paper proposes OFFLOAD, an open-source, hardware-agnostic software-hardware framework. OFFLOAD facilitates the distribution of tasks across diverse hardware units, encompassing both cutting-edge accelerators and existing system-on-chip (SoC) architectures. Our framework seamlessly integrates with popular databases and application development tools. Through the utilization of multi-level abstractions implemented at the compiler, operating system, and driver levels, OFFLOAD translates high-level code and data into hardware-optimized binary instructions. To the best of our knowledge, OFFLOAD represents a ground-breaking approach within this domain. The feasibility of OFFLOAD is demonstrably validated by its integration with prevalent tools such as MySQL, Apache Spark, and Apache Arrow within a user-friendly Python environment. Subsequently, tasks are offloaded for execution on hardware leveraging memory mapped I/O. This is exemplified by integrating OFFLOAD with Raspberry Pi devices, showcasing the entire workflow from software-based data query to hardware execution.
Table of Contents
Introduction -- Flow of information -- Software stack -- Hardware setup -- Hardware network and task distribution -- Software & firmware implementation -- Results -- Conclusion
DOI
PubMed ID
Degree
M.S. (Master of Science)