Multi–Task Scheduling For PIM-Based Heterogeneous Computing System

With the advent of big data era,various data-intensive applications have undergone a spurt-style development,which has made the problem of "memory wall" more and more acute.Process in Memory is a new computing architecture that can effectively solve the "memory wall" problem.The main principle of the PIM structure is to tightly couple the computing unit and the storage unit so as to eliminate the limitation of the memory bandwidth bottleneck and the overhead caused by the data migration between the computing unit and the storage unit.The computing system containing PIM structure is a new type of heterogeneous computing architecture and has been studied continuously this year.However,the task scheduling model suitable for PIM + CPU heterogeneous systems still lacks corresponding attention and research.This paper studies the PIM + CPU heterogeneous system from two aspects.Firstly,aiming at the high-speed and low-latency memory access characteristics of PIM architecture,a task scheduling model for PIM + CPU heterogeneous system is established.At the same time,a task scheduling algorithm,PIM-FAST,is proposed to optimize system performance.Secondly,aiming at the aging problem of PIM + CPU heterogeneous system,a task scheduling algorithm for mitigating system aging(PIM-MTTF)is proposed from the operating system level to improve the reliability of the system.In order to enhance the performance of PIM + CPU heterogeneous systems,this paper considers the existence of memory access competition in multi-task parallel execution,and establishes a task-based quantization model and corresponding task scheduling algorithm for PIM + CPU heterogeneous systems.In this model,an application is split into many tasks,and each task is mapped to a suitable processing unit(PIM or per CPU core)for execution according to the processing unit mapping algorithm PIM-FAST.Considering the memory competition,PIM-FAST can improve the performance of PIM + CPU heterogeneous systems.Experimental data show that PIM-FAST can bring more performance improvement than HEFT,which is a task scheduling algorithm for traditional heterogeneous systems.In order to improve the reliability of PIM + CPU heterogeneous system,this paper also proposes a reliability-aware task scheduling algorithm(PIM-MTTF),for PIM + CPU heterogeneous systems based on MTTF(Mean Time To Failure)model.PIM-MTTF achieves the goal of reducing the difference in reliability between different processing units in PIM + CPU heterogeneous system.As can be seen from the experimental data,compared with the traditional task scheduling algorithm to enhance system reliability,PIM-MTTF algorithm is more suitable for improving the reliability of the PIM + CPU heterogeneous system.