AI Cluster Data Caching via Weighted Cache Path Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The storage resources in AI clusters act as a bottleneck, affecting data cache performance due to high storage requirements and frequent input/output operations.
Innovation Solution
A method for caching data in AI clusters involves determining a target data set, acquiring weight values on each node, selecting a target cluster node, and establishing a cache path using the weight values and shortest paths to optimize data distribution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data is cached in AI cluster storage resources, then data availability is improved, but storage resources become a bottleneck due to high storage requirements and frequent I/O operations
Solution Approach 1:
The patent segments the AI cluster into multiple node types (management nodes, shared storage nodes, and non-management nodes) with differentiated caching functions. Management nodes determine caching strategies, shared storage nodes provide centralized storage, and non-management nodes perform local caching, thereby distributing I/O operations and avoiding storage bottlenecks while maintaining data availability.
Solution Approach 2:
The patent introduces management nodes as intermediaries that coordinate between shared storage nodes and non-management nodes. These management nodes handle caching strategy determination and data distribution, reducing direct I/O pressure on shared storage nodes while ensuring data availability through coordinated caching across the cluster.
2Speed
If data is cached locally on cluster nodes, then data access speed is improved, but storage resource distribution becomes unbalanced
Solution Approach 1:
The patent implements local quality by assigning different caching roles to different node types based on their characteristics. Non-management nodes with sufficient local storage perform local caching for faster access, while management nodes with higher computational capability perform caching strategy determination. This differentiated approach optimizes data access speed for each node type while maintaining balanced resource distribution through centralized coordination.
3Productivity
If caching strategy is optimized for specific nodes, then data cache performance is improved, but system complexity increases
Solution Approach 1:
The patent achieves universality by creating a multi-functional caching system where management nodes perform both coordination functions (determining caching strategies for other nodes) and local caching functions. Shared storage nodes provide both centralized storage and data distribution functions. This multi-functionality allows the system to optimize cache performance across different node types while using a unified management framework that prevents excessive system complexity.
Data Source
AI summary
The present application discloses a method for caching data in AI cluster, system, apparatus and computer medium. The method comprises: determining target data set to be cached; acquiring weight value of the target data set on each of a plurality of cluster nodes of the AI cluster; determining a target cluster node for caching the target data set; acquiring a target shortest path from remaining cluster nodes of the AI cluster to the target cluster node, and a predecessor node of the target cluster node in the target shortest path, wherein the remaining cluster nodes comprise nodes of the AI cluster except the target cluster node; determining, based on the weight value, the target shortest path and the predecessor node, a cache path for caching the target data set to the target cluster node, to cache the target data set to the target cluster node according to the cache path.