Dynamic Virtual GPU Scheduling for Heterogeneous Resource Allocation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional GPU virtualization technologies restrict dynamic adjustment of virtual GPU performance and resource allocation, allowing only homogeneous virtual GPUs to run on a physical GPU, leading to inefficient resource utilization and inability to support heterogeneous virtual GPUs.
Innovation Solution
A dynamic scheduling system that allocates time slices and frame buffer space to virtual GPUs, enabling the creation, deletion, and upgrading of virtual GPUs based on resource requirements, allowing multiple heterogeneous virtual GPUs to run on a single physical GPU.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If a physical GPU is divided into multiple virtual GPUs with fixed specifications, then resource allocation is simplified and management is easier, but the system cannot dynamically adjust virtual GPU performance and cannot support heterogeneous virtual GPUs
Solution Approach 1:
The patent implements dynamic virtual GPU specification adjustment by allowing the virtualization management module to modify virtual GPU configurations at runtime. The system can dynamically change the number of virtual GPUs, their specifications, and resource allocation without requiring physical reconfiguration, thus achieving both ease of management and specification flexibility.
Solution Approach 2:
The system changes key parameters of virtual GPUs dynamically, including the number of virtual GPUs, their performance levels (e.g., 1/2, 1/4, 1/8 of physical GPU performance), and resource allocation ratios. This allows the same physical GPU to support multiple heterogeneous virtual GPU types simultaneously, resolving the contradiction between fixed management and flexible adaptation.
2Stability of the object's composition
If virtual GPUs are statically configured at cluster deployment, then system stability is improved and data consistency is maintained, but resource utilization efficiency decreases and dynamic adjustment is impossible
Solution Approach 1:
The patent introduces dynamic configuration capabilities that allow virtual GPU specifications and resource allocation to be adjusted at runtime while maintaining system stability. The virtualization management module can respond to changing workload demands, creating or deleting virtual GPUs and adjusting their specifications dynamically, thus improving resource utilization without sacrificing configuration consistency.
Solution Approach 2:
The system implements feedback mechanisms where the virtualization management module monitors resource usage and workload demands, then dynamically adjusts virtual GPU configurations accordingly. This feedback loop enables the system to maintain stability while adapting to changing conditions, optimizing resource utilization efficiency without compromising data consistency.
3Device complexity
If only homogeneous virtual GPUs run on a physical GPU, then system complexity is reduced and management is simplified, but resource utilization efficiency decreases and flexibility is limited
Solution Approach 1:
The patent makes the physical GPU universally compatible with multiple types of virtual GPUs by implementing a virtualization layer that can create and manage heterogeneous virtual GPU instances. The same physical GPU can simultaneously host virtual GPUs with different specifications (e.g., high-performance, medium-performance, low-performance types), allowing diverse workloads to be accommodated without increasing physical hardware complexity.
Solution Approach 2:
The system dynamically changes virtual GPU specification parameters to match different workload requirements. The virtualization management module can allocate different performance levels and resource ratios to various virtual GPUs running on the same physical GPU, enabling efficient resource utilization across homogeneous and heterogeneous virtual GPU types without complicating the underlying hardware architecture.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Embodiments of the present application provide a method for implementing a virtual GPU. The method for implementing a virtual GPU includes: allocating to each of the virtual GPUs a running time slice corresponding to the resource requirement of the virtual GPU according to resource requirements of virtual GPUs running on the same physical GPU, wherein a sum of running time slices of all virtual GPUs configured on a physical GPU is less than or equal to a scheduling period; and allocating resources of the physical GPU to the virtual GPUs according to the running time slices allocated to the virtual GPUs.