Cloud FPGA Graph Acceleration for GNN Memory Access
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems face inefficiencies in accessing and processing graph data due to random memory accesses, which can lead to high costs and reduced performance in graph neural network (GNN) processing.
Innovation Solution
A system comprising a host with a circuitry board that includes an access engine circuitry, which fetches structure and attribute data from pinned memory via PCIe connections, performs node sampling, and sends the data to processors for GNN processing, thereby optimizing memory allocation and access for sparse matrix multiplications.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If graph data is stored in memory for GNN processing, then data availability is improved, but random memory access patterns cause high costs and reduced performance
Solution Approach 1:
The access engine performs preliminary actions by fetching and preprocessing graph data (structure data and attribute data) before the actual GNN processing occurs. It proactively retrieves data from memory, performs node sampling to identify relevant nodes, and prepares the data in an optimized format, thereby eliminating the need for random access patterns during the actual processing phase and improving overall system performance
Solution Approach 2:
The access engine acts as an intermediary component between the memory system and the GNN processing units. It mediates the data flow by receiving requests from the processing system, fetching data from memory, performing node sampling to filter relevant data, and delivering the processed data to the GNN processing units. This intermediary role transforms random access patterns into structured, efficient data retrieval operations
2Reliability
If all graph data is fetched for processing, then completeness is improved, but data transmission volume and processing time increase
Solution Approach 1:
The access engine extracts only the necessary portion of graph data by performing node sampling. Instead of fetching and processing all graph data, it identifies and extracts the subset of data corresponding to sampled nodes that are relevant for the specific GNN processing task. This extraction principle reduces data transmission volume and processing time while maintaining the completeness of the required data subset
Solution Approach 2:
The system applies partial action by fetching and processing only a portion of the graph data corresponding to sampled nodes rather than the entire graph. The node sampling mechanism determines the appropriate subset of nodes and their associated attribute data, enabling the system to perform GNN processing with partial data that is sufficient for the task at hand, thereby reducing processing time and resource consumption
Data Source
AI summary
This application describes systems and methods for facilitating memory access for graph neural network (GNN) processing. An example method includes fetching, by an access engine circuitry implemented on a circuitry board, a portion of structure data of a graph from a pinned memory in a host memory of a host via a first peripheral component interconnect express (PCIe) connection; performing node sampling using the fetched portion of the structure data of the graph to select one or more sampled nodes; fetching, by the access engine circuitry, a portion of attribute data of the graph from the pinned memory via the first PCIe connection; sending the fetched portion of the attribute data of the graph to one or more processors; and performing, by the one or more processors, GNN processing for the graph using the fetched portion of the attribute data of the graph.