Tensor Data Parallel Processing Conflict Resolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for analyzing multidimensional tensor data, such as canonical polyadic decomposition (CPD) using alternating least squares (ALS), face significant challenges with increased calculation time and memory usage due to the matricized tensor times khatri-rao product (MTTKRP) operations, particularly when dealing with sparse data, where conflicts between parallel operations can lead to incorrect results.
Innovation Solution
An information processing method that acquires data specifying non-zero elements and their indices, generates data for grouping overlapping index combinations into different groups, and performs MTTKRP processing in parallel on these groups to avoid conflicts and optimize calculation time and memory usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If MTTKRP operations are performed in parallel on sparse tensor data, then calculation time is reduced, but conflicts between parallel operations occur leading to incorrect results
Solution Approach 1:
The patent segments the parallel MTTKRP operations into multiple independent groups based on index patterns. By dividing the computation space into discrete segments that do not overlap in memory access, the patent eliminates conflicts between parallel threads while maintaining parallelization benefits. Each segment processes a unique subset of tensor elements, ensuring correctness while achieving speedup.
2Productivity
If conventional MTTKRP methods are used for tensor decomposition, then analysis of multidimensional data is achieved, but memory usage increases significantly
Solution Approach 1:
The patent extracts and processes only the non-zero elements of sparse tensor data, ignoring zero elements entirely. By extracting the essential computational information (non-zero values and their indices) and performing MTTKRP operations only on these extracted elements, the patent dramatically reduces memory usage while preserving the full analytical capability for tensor decomposition.
3Productivity
If parallel processing is applied to MTTKRP operations, then calculation efficiency is improved, but device complexity increases due to conflict management
Solution Approach 1:
The patent performs preliminary grouping of computation tasks before parallel execution. By pre-organizing the MTTKRP operations into conflict-free groups based on index patterns, the patent eliminates the need for complex runtime conflict detection and resolution mechanisms. This preliminary organization simplifies the parallel processing architecture while maintaining high calculation efficiency.
Data Source
AI summary
A non-transitory computer-readable recording medium stores an information processing program for causing a computer to execute processing including: acquiring first data that enables, for each of non-zero elements included in multidimensional tensor data, specification of a combination of a value of the element and an index of each dimension that indicates a position of the element; generating, on the basis of the acquired first data, second data that enables specification of a plurality of groups obtained by grouping each of the combinations such that the combinations with indexes that overlap with each other are included in different groups; and performing, on the basis of the generated second data, matricized tensor times khatri-rao product (MTTKRP) processing by setting each combination of a plurality of combinations included in the group as a target of parallel processing in the MTTKRP processing related to the tensor data.