Computational SSD Accelerating Graph Neural Network Preprocessing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Graph neural network (GNN) machine learning processes face bottlenecks and memory limitations during preprocessing, particularly in loading graph data from storage and sampling, which hinder practical system application.
Innovation Solution
A computational solid state drive (SSD) with a field-programmable gate array (FPGA) is used to accelerate both GNN inference and preprocessing by converting graph machine learning models into data flow graphs, allowing direct access and execution, and generating a graph machine learning inference accelerator, with a programmable operation unit and separate areas for fixed and dynamic hardware logic.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If graph data is loaded from storage to memory for GNN preprocessing and inference, then the GNN process can be executed, but serious bottlenecks and memory limitations occur
Solution Approach 1:
The patent combines storage and computation functions into a single computational storage device. The SSD includes both storage units for graph data and execution units for GNN operations, eliminating the need to transfer large-scale graph data between storage and memory systems. This merging resolves the memory limitation bottleneck by enabling direct processing of graph data at the storage level.
Solution Approach 2:
The patent introduces an intermediate buffer memory within the computational storage device that acts as a mediator between the storage units and execution units. This buffer memory accommodates the graph data during preprocessing and inference operations, eliminating the need for external memory systems and resolving the memory capacity limitation while maintaining high processing speed.
2Adaptability or versatility
If existing neural network-based machine learning systems are used for GNN inference, then inference can be performed, but preprocessing bottlenecks and memory limitations prevent practical system application
Solution Approach 1:
The computational storage device is designed with multi-functional execution units that can perform both GNN preprocessing operations (such as sampling and feature extraction) and GNN inference operations. This universal design enables the single device to handle the entire GNN workflow, improving adaptability while simplifying system operation by eliminating the need for separate preprocessing and inference systems.
3Productivity
If large-scale graph data is moved between storage and memory systems, then GNN processing can be performed, but serious bottlenecks occur
Solution Approach 1:
By merging storage and computation functions into a single computational storage device, the patent eliminates the data transfer step between separate storage and memory systems. The execution units are directly integrated with the storage units, allowing graph data to be processed in-place without movement, thereby eliminating transfer time bottlenecks and improving overall processing throughput.
Data Source
AI summary
A computational storage supporting graph machine learning acceleration includes a solid state drive (SSD) configured to store a graph data set; and a field-programmable gate array (FPGA) configured to download, to a memory, a graph machine learning model programmed in a form of a data flow graph by a host, wherein a hardware logic built in the FPGA performs access to the SSD through a peripheral component interconnect-express (PCIe) switch.