GPU Frequent Pattern Mining via Block Partitioning and Memory Management
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing frequent itemset mining methods face challenges with slow performance, limited scalability, and high computational intensity due to the limitations of single-thread CPU processing and the small GPU memory, leading to inefficiencies in handling large datasets and workload skewness.
Innovation Solution
The proposed method involves generating blocks from bit vectors corresponding to frequent 1-itemsets, copying relative memory addresses of candidate k-itemsets to GPU device memories, and updating their supports by synchronizing partial calculations, utilizing a traversal from the first level (TFL) strategy to reduce memory usage and improve performance, and employing a hopping from intermediate level (HIL) strategy to handle long patterns efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If GPU-based methods are used to accelerate frequent pattern mining, then computational performance is improved, but the limited GPU memory capacity restricts the size of data that can be processed
Solution Approach 1:
The patent divides the large dataset into multiple blocks that can be processed in parallel across multiple GPUs. Each GPU handles a specific block of data, enabling the system to process datasets larger than the memory capacity of individual GPUs while maintaining high computational performance through parallel processing.
Solution Approach 2:
The patent introduces a distributed memory architecture across multiple GPUs, transitioning from a single-GPU memory constraint to a multi-GPU collective memory space. This dimensional expansion allows the system to handle larger datasets by distributing data across multiple memory devices while maintaining efficient access patterns through the proposed indexing mechanism.
2Quantity of substance
If multiple GPUs are used to process larger datasets, then data processing capacity is improved, but workload skewness reduces efficiency
Solution Approach 1:
The patent implements a localized indexing mechanism where each GPU maintains an index for its assigned data block. This local quality approach ensures that each GPU can independently and efficiently access and process its local data without requiring synchronized access to global data structures, thereby eliminating workload skewness and improving parallel processing efficiency.
Solution Approach 2:
The patent performs preliminary data partitioning and index construction before the actual frequent pattern mining process. By pre-organizing data into balanced blocks with local indexes, the system ensures that workload is evenly distributed across multiple GPUs from the outset, preventing workload skewness during the mining process.
3Speed
If intermediate data is kept at GPU memory level for fast access, then processing speed is improved, but memory capacity limits the amount of intermediate data that can be stored
Solution Approach 1:
The patent introduces a hierarchical memory structure with GPU device memory as the fast intermediate storage and host memory as the bulk storage. Intermediate data that exceeds GPU memory capacity is automatically spilled to host memory through a managed memory mechanism, allowing the system to maintain fast access patterns for frequently used data while accommodating large volumes of intermediate data in the slower but larger host memory.
Data Source
AI summary
Disclosed is a frequent pattern mining method and apparatus, the frequent pattern mining apparatus that may copy relative memory addresses of candidate itemsets, from a main memory to device memories of graphic processing units (GPUs), copy at least one same block required for calculating supports of the candidate itemsets, from the main memory to the device memories, and update the supports of the candidate itemsets by synchronizing partial supports processed by the GPUs.