SAGE: A Compiler-assisted Reinforcement Learning-based Offloading Approach under Near-memory Processing Paradigm

Emerging data-intensive applications face significant performance challenges due to huge time and energy costs associated with off-chip data transmission under the traditional Von Neumann architecture. The Near-Memory Processing (NMP) paradigm offers a solution by enabling computation within or near the memory. This method enhances system performance by offloading some portions of the application to the memory side, thereby reducing the need for extensive off-chip data movement and mitigating the associated inefficiencies. Although several static and dynamic offloading approaches have been proposed recently, they often lack the adaptability to accommodate unknown application types and dynamic system conditions. Therefore, this article presents an adaptive hybrid computation offloading approach, SAGE, which leverages various compiler-extracted program information and employs a reinforcement learning module that dynamically adapts the system’s condition to make offloading decisions. Moreover, this approach considers a basic block-level offloading granularity, which is neither fine- nor coarse-level granularity, to minimize the overhead and enable a dynamic performance-controlled offloading mechanism. We validate the proposed approach, SAGE, through experiments with various types of applications, and the results demonstrate its effectiveness across several aspects. SAGE saves execution time by 24% (overall) and reduces energy consumption by 19% (overall) compared to the traditional computing approach, in which computation occurs only on the host processor. Furthermore, compared to state-of-the-art approaches, SAGE saves the execution time by 58% (maximum) and reduces the overall energy consumption by 36% (maximum). Furthermore, SAGE reduces the off-chip data transfer by 52% (overall) compared to the traditional computing approach.