Unified Memory Architecture for CPU-GPU Task Scheduling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computing systems face challenges in managing task scheduling between CPUs and GPUs, including separate memory systems, quality of service guarantees, programming models, and efficient scheduling, which lead to inefficiencies and power consumption issues, especially in heterogeneous processing environments.
Innovation Solution
A unified computing system architecture that includes a CPU and an Accelerated Processing Device (APD) with a software scheduler and kernel mode driver, allowing for fluid task switching and resource management between CPU and GPU, enabling efficient scheduling and resource utilization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If separate memory systems are used for CPU and GPU, then each processor has dedicated memory access, but memory latency and power consumption increase due to chip-to-chip interfaces
Solution Approach 1:
The patent merges the memory systems of the CPU and GPU into a unified memory architecture, eliminating the need for separate memory systems and chip-to-chip interfaces. This consolidation reduces power consumption while maintaining dedicated memory access capabilities for both processors through shared memory resources.
2Reliability
If separate memory systems are used for CPU and GPU, then each processor has dedicated memory access, but memory overhead becomes unacceptable for fine grain offload
Solution Approach 1:
The patent combines separate memory systems into a unified memory architecture, reducing memory overhead and simplifying the system. This enables fine-grain offload operations by eliminating redundant memory structures and interfaces, while both CPU and GPU can access the shared memory efficiently.
3Ease of operation
If inefficient scheduling is used in multiple processing device systems, then process monitoring is required, but GPU hardware cannot be easily identified and preempted
Solution Approach 1:
The patent implements a scheduling mechanism where the system automatically identifies, monitors, and preempts GPU processes without requiring external software intervention. The GPU hardware itself participates in the scheduling decisions, enabling efficient task switching while maintaining ease of process monitoring through built-in scheduling capabilities.
4Adaptability or versatility
If software monitoring of GPU utilization is implemented, then resource management is possible, but the software burden increases
Solution Approach 1:
The patent enables the GPU system to self-manage its own resource utilization and scheduling without requiring heavy software intervention. The hardware includes built-in mechanisms for tracking GPU usage, identifying processes, and making scheduling decisions, thereby reducing the software burden while maintaining adaptability for resource management.
Data Source
AI summary
Embodiments describe herein provide a method of for managing task scheduling on a accelerated processing device. The method includes executing a first task within the accelerated processing device (APD), monitoring for an interruption of the execution of the first task, and switching to a second task when an interruption is detected.