GPU Resource Management via Virtual Mapping Logic
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current GPU-based application programs require redesigning to accommodate multiple GPUs, leading to increased development costs and inefficiencies, as they need to account for low-level hardware details such as video memory capacities and computing units.
Innovation Solution
A resource management method and apparatus that maps available GPU hardware resources to a unified virtual GPU resource, allowing application programs to allocate resources without needing to redesign for specific hardware structures, thereby separating the application from low-level GPU details and enabling efficient use of GPU resources across multiple GPUs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If application programs are redesigned to accommodate multiple GPUs and low-level hardware details, then GPU resource utilization efficiency is improved, but application development cost and complexity increase
Solution Approach 1:
The patent introduces a resource management device as an intermediary layer between the application program and the physical GPUs. This device maintains mapping logic that translates unified virtual GPU resource requests into specific allocations across multiple physical GPUs, allowing applications to interact with a simplified interface while achieving efficient multi-GPU resource utilization underneath
Solution Approach 2:
The resource management device provides a unified virtual GPU resource interface that can represent and manage multiple different physical GPU configurations. The same application program can run on various hardware structures (single GPU, multiple GPUs, different video memory capacities) without modification, as the resource management device adapts the unified interface to the specific underlying hardware
2Productivity
If application programs are designed for specific hardware structures, then GPU resource efficiency is improved, but portability across different hardware structures deteriorates
Solution Approach 1:
The resource management device creates a universal interface layer that can map to different underlying hardware configurations. The unified virtual GPU resource can represent single GPU, multiple GPUs, or combinations with different video memory capacities, allowing the same application to achieve efficient resource utilization across diverse hardware structures without redesign
Solution Approach 2:
The patent segments the hardware abstraction into two layers: a unified virtual GPU resource layer that applications interact with, and the underlying physical GPU layer with specific hardware details. The resource management device maintains mapping logic between these layers, allowing the application layer to remain hardware-agnostic while the lower layer handles hardware-specific optimizations
3Power
If multiple GPUs are used to increase computational capacity, then computing power is improved, but the need for redesigning application programs increases
Solution Approach 1:
The resource management device acts as a mediator that handles the complexity of coordinating multiple GPUs. It receives resource allocation requests for the unified virtual GPU resource and distributes them across multiple physical GPUs using maintained mapping logic, thereby enabling scalable computational capacity without requiring applications to be redesigned for multi-GPU architectures
Data Source
AI summary
Embodiments of the present invention provide resource managing methods and systems. The method comprises: receiving a request to allocate resources sent from host code of an application program located on a first device; in accordance with the allocation request and a maintained mapping logic mapping available hardware resources of at least one graphics processing unit (GPU) of the first device to a unified virtual GPU resource, allocating required resources for a device code of the application program from the available hardware resources of at least one GPU of the first device; and forwarding information of the allocated resource back to the host code. The present invention can efficiently utilize GPU resources and reduce implementation costs.