Automated Parallelization Kernel Selection for Distributed Simulations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Selecting the optimal parallelization method for distributed simulations is computationally prohibitive and requires a priori evaluation, which can be time-consuming and inefficient, as it ties the parallelization method and model simulation together, making it difficult to achieve efficient memory consumption and computation times.
Innovation Solution
A method for automating parallelization kernel selection in distributed simulations by performing simulation runs with both external and internal kernels, determining successful runs, and generating output data, allowing for dynamic selection based on run results without initial optimization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a proper parallelization method is chosen in advance for a particular model, then memory consumption and run time are significantly reduced, but the estimation of optimality is computationally prohibitive and can take as much or more time as running the simulation itself
Solution Approach 1:
The patent performs preliminary simulation runs with a default parallelization method to gather performance data before selecting the optimal method. This preliminary action provides the basis for subsequent optimization without requiring a full a priori estimation, thus reducing the time loss while still enabling productivity improvements.
Solution Approach 2:
The patent implements dynamic selection of parallelization methods based on runtime performance monitoring. Instead of static a priori selection, the system dynamically adjusts the parallelization method based on actual simulation performance, eliminating the need for time-consuming preliminary estimation while maintaining optimization benefits.
2Productivity
If parallelization method selection is tied to model simulation, then greater efficiencies are realized, but the process becomes complex and requires choosing the method in advance
Solution Approach 1:
The patent enables the simulation system to automatically select the optimal parallelization method based on runtime performance data without requiring external intervention or complex pre-planning. The system serves itself by monitoring performance and autonomously adjusting parallelization settings, thus maintaining efficiency while reducing process complexity.
Solution Approach 2:
The patent implements a feedback mechanism where simulation performance data is continuously monitored and used to inform parallelization method selection. This closed-loop approach simplifies the selection process by using actual performance feedback rather than requiring complex a priori analysis, thereby maintaining productivity while reducing complexity.
Data Source
AI summary
Systems and methods of automating parallelization kernel selection in distributed simulations are discussed herein. Techniques disclosed pertain to attempting execution of all runs or representative run sets on an external simulation kernel and monitoring for failures, with a threshold amount of failures prompting simulation run completion on an internal simulation kernel. Further techniques pertain to performing static analysis on a model simulation job to determine an appropriate simulation kernel.