GPU Cluster Task Mapping by Interconnect Hierarchy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In GPU clusters, inferior communication lines lead to reduced parallel efficiency and performance issues due to varying communication speeds between GPUs, hindering data transmission and computation.
Innovation Solution
A method to obtain a graph of interconnections between GPUs, form a parallel hierarchical architecture based on these connections, and map parallel tasks to optimize task assignment, ensuring high-speed connections are used for tasks with high communication requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If tasks are assigned to multiple GPUs without considering interconnection quality, then parallel computation capability is improved, but communication efficiency deteriorates due to inferior communication lines
Solution Approach 1:
The patent applies local quality by differentiating communication channels based on their interconnection quality. It identifies high-speed communication lines (e.g., NVLink, PCIe) versus low-speed lines (e.g., InfiniBand, Ethernet) and routes communication tasks through appropriate channels. This ensures that critical data transmissions use optimal paths, resolving the contradiction between parallel computation and communication efficiency.
Solution Approach 2:
The patent implements dynamic task assignment that adapts to the GPU cluster's interconnection topology. The system dynamically selects which GPUs to assign tasks to based on real-time communication requirements and interconnection quality, rather than using static assignment. This dynamic approach allows the system to maintain high parallel computation capability while avoiding communication bottlenecks.
2Device complexity
If inferior communication lines are selected for GPU interconnection, then device complexity is reduced, but parallel efficiency deteriorates due to communication bottlenecks
Solution Approach 1:
The patent segments the GPU cluster into sub-groups based on interconnection quality, creating a hierarchical structure. High-speed interconnected GPUs form tight sub-groups for compute-intensive tasks, while low-speed connections are used for less critical communications. This segmentation allows the system to maintain simplicity in individual sub-groups while achieving high overall parallel efficiency through coordinated operation.
Solution Approach 2:
The patent introduces a task assignment mechanism that acts as an intermediary between computation tasks and GPU resources. This intermediary evaluates interconnection quality and makes intelligent routing decisions, allowing inferior communication lines to be used only when necessary while directing critical communications through high-speed paths, thereby maintaining parallel efficiency without requiring complete high-speed interconnection.
Data Source
AI summary
Illustrative embodiments of the present disclosure include a method, a device, and a product for a Graphics Processing Unit (GPU) cluster. The method includes: obtaining a graph of interconnections between GPUs in the GPU cluster; forming a parallel hierarchical architecture of the GPU cluster based on the graph of interconnections between GPUs; and mapping parallel tasks to the parallel hierarchical architecture to execute the parallel tasks. The method for a GPU cluster according to the present disclosure ensures that high-speed GPU-GPU connection is used for tasks with high communication requirements, thus improving the overall processing efficiency of the GPU cluster.