Hierarchical Imitation Learning Scheduler for Many-Core Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing optimization-based techniques for runtime task scheduling in heterogeneous many-core systems are inefficient due to the combinatorial nature of the task scheduling problem, leading to suboptimal performance and high computational overhead, especially in dynamic environments with multiple overlapping tasks.
Innovation Solution
A hierarchical imitation learning-based scheduler that learns from an oracle to maximize performance and energy efficiency by posing task scheduling as a classification problem, using a combination of static and dynamic features to predict optimal task assignments to processing elements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If optimization-based techniques (MIP, CP) are used for runtime task scheduling, then scheduling quality is improved, but computational overhead becomes prohibitive
Solution Approach 1:
The patent pre-computes and stores optimal scheduling decisions in lookup tables during system initialization or offline phases. During runtime, the scheduler simply retrieves pre-computed decisions based on current system state, avoiding expensive optimization computations and achieving both high scheduling quality and low computational overhead.
Solution Approach 2:
The patent creates simplified models or approximations of the complex scheduling problem that can be solved quickly. These models capture the essential characteristics of optimal scheduling without requiring full optimization computations, enabling fast runtime decisions that closely approximate optimal solutions.
2Ease of manufacture
If static task-to-PE mapping is used, then implementation simplicity is improved, but parallel instance utilization deteriorates
Solution Approach 1:
The patent implements dynamic task-to-PE mapping where the scheduler continuously monitors system state and adjusts task assignments in real-time. This allows the system to adapt to changing conditions, maximize parallel instance utilization, and maintain high productivity while keeping the implementation manageable through modular design.
Solution Approach 2:
The scheduler automatically makes intelligent decisions about task placement without requiring manual configuration or complex external control. It self-adapts to system conditions and optimizes resource utilization autonomously, achieving both simplicity and high performance.
3Loss of time
If heuristic schedulers (SJF, CFS) are used, then computational overhead is reduced, but scheduling decision quality deteriorates
Solution Approach 1:
The patent uses pre-computed optimal scheduling decisions stored in lookup tables that capture the essence of high-quality scheduling. During runtime, simple table lookups provide both fast execution and high scheduling decision quality, effectively copying the benefits of complex optimization without the computational cost.
Solution Approach 2:
The patent transforms the scheduling problem from requiring complex real-time optimization to using pre-computed solutions with simple runtime parameter matching. By changing the approach from computation-intensive to data-retrieval-intensive, it achieves both low overhead and high decision quality.
4Productivity
If domain-specific architectures are used, then performance and energy efficiency are improved, but programming flexibility deteriorates
Solution Approach 1:
The patent creates a unified scheduling framework that works across diverse domain-specific architectures and application types. The scheduler is designed to be architecture-agnostic and can efficiently manage tasks on different types of processing elements, maintaining programming flexibility while achieving high performance on domain-specific hardware.
Solution Approach 2:
The patent separates the scheduling decision-making logic from the hardware architecture details. By creating an independent scheduling layer that abstracts away hardware specifics, it enables domain-specific architectures to maintain their performance advantages while preserving programming flexibility through a universal scheduling interface.
Data Source
AI summary
Runtime task scheduling using imitation learning (IL) for heterogenous many-core systems is provided. Domain-specific systems-on-chip (DSSoCs) are recognized as a key approach to narrow down the performance and energy-efficiency gap between custom hardware accelerators and programmable processors. Reaching the full potential of these architectures depends critically on optimally scheduling the applications to available resources at runtime. Existing optimization-based techniques cannot achieve this objective at runtime due to the combinatorial nature of the task scheduling problem. In an exemplary aspect described herein, scheduling is posed as a classification problem, and embodiments propose a hierarchical IL-based scheduler that learns from an Oracle to maximize the performance of multiple domain-specific applications. Extensive evaluations show that the proposed IL-based scheduler approximates an offline Oracle policy with more than 99% accuracy for performance- and energy-based optimization objectives. Furthermore, it achieves almost identical performance to the Oracle with a low runtime overhead and high adaptivity.