Graph-Processing Accelerator for Scalable Distributed Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing graph computing systems face performance bottlenecks and lack flexibility due to the large size of graph datasets, which are often distributed across multiple memory devices, leading to increased communication latency and varying hardware and software interfaces when scaling.
Innovation Solution
A scalable graph-processing accelerator architecture that deploys specialized hardware units such as SIMD or systolic-array-based throughput processing units, vertex processing units, format shuffle units, and programmable gather/scatter units, allowing for graph data processing close to where the data resides, with a unified programming interface across different locations in the computing system.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If graph datasets are distributed over multiple memory devices to handle large-scale data, then the system can process larger graph datasets, but communication latency increases and performance bottlenecks occur
Solution Approach 1:
The system segments the graph processing workload across multiple computing nodes, with each node handling a specific partition of the graph dataset. This segmentation allows parallel processing while maintaining data locality, reducing the need for frequent inter-node communication and mitigating latency issues associated with distributed memory access.
Solution Approach 2:
The patent introduces a unified programming interface that acts as an intermediary between the user and the distributed graph processing system. This interface abstracts the complexity of data distribution and communication, providing a consistent API that handles data placement, retrieval, and processing across multiple memory devices without exposing communication latency to the user.
2Quantity of substance
If multiple computing nodes are used to scale the graph computing system, then larger graph datasets can be processed, but performance bottlenecks and lack of flexibility occur
Solution Approach 1:
The patent implements a universal programming interface that provides consistent functionality across different computing nodes and configurations. This interface allows the same graph processing algorithms to be executed on various hardware platforms (CPU, GPU, FPGA) without modification, enhancing system flexibility and adaptability while maintaining the ability to scale across multiple nodes.
Solution Approach 2:
The system employs dynamic data placement and processing strategies that can adapt to different hardware configurations and workload characteristics. The unified interface allows runtime adjustments in data distribution and processing allocation across computing nodes, enabling the system to optimize performance and maintain flexibility as the scale and requirements change.
3Productivity
If specialized hardware units are deployed for graph processing, then processing performance is improved, but hardware complexity increases
Solution Approach 1:
The patent implements self-service mechanisms where the unified programming interface automatically manages data formatting, shuffling, and gathering operations. The system includes built-in format shuffle units and gather/scatter units that autonomously handle data transformation between different storage formats and processing formats, reducing the burden on users and simplifying the effective hardware complexity they must manage.
Solution Approach 2:
The unified programming interface serves as an intermediary layer between the user and the complex specialized hardware units. It abstracts the complexity of SIMD processors, systolic arrays, and other specialized components by providing a consistent high-level interface, allowing users to leverage hardware acceleration without directly managing the underlying hardware complexity.
Data Source
AI summary
An accelerator device includes a first processing unit to access a structure of a graph dataset, and a second processing unit coupled with the first processing unit to perform computations based on data values in the graph dataset.