The Research Of Task Partitioning And Scheduling For Novel PIM Heterogeneous System

Processing-in-Memory (PIM) is returning as a promising solution to address the issue of memory wall as computing systems gradually step into the big data era. The core concept of PIM or Near-Data-Computing (NDC) is to closely couple memory/storage components and computation resources to get rid of the bandwidth bottleneck and communication cost between CPU and memory. The system that consists of PIM engine and host processor is a novel heterogeneous parallel computing architecture. Researchers continually proposed various PIM architecture combined with novel memory device or 3D integration technology, but it is still a lack of universal task scheduling method in terms of the new heterogeneous platform.Considering the customized PIM architectures are limited by redundant hardware and the lack of generality, we propose a formalized model to quantify the performance and energy of the PIM+CPU heterogeneous parallel system based on task-graph analysis, and also propose a PIM-aware task partitioning and mapping framework to exploit different PIM engines. In this framework, an application is divided into subtasks and each of the subtasks is mapped onto appropriate execution units (CPU or PIM)according to the proposed PIM-oriented Earliest-Finish-Time (PEFT) algorithm, which parallelizes the execution of subtasks both on host processor and computational DRAM to maximize the performance gains brought by PIM. Based on the optimized scheduling sequence of subtasks, The PEFT algorithm selects the unit which can minimize the execution finish time for each subtask. The data-intensive machine learning applications are selected as benchmarks for the off-the-shelf 3D-stacked DRAM product, HMC-2.0 module. Experimental evaluations show our task partitioning and mapping framework'reduces the execution time of benchmarks by 46% compared to conventional computing architectures, so as to significantly improve the system performance.