Tree-Based Space-Partitioning for Heterogeneous Load Balancing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computer simulation methods face challenges in load balancing across heterogeneous computer clusters, particularly in complex seismic modeling, due to varying computational capacities and changing resource availability, leading to inefficiencies and increased costs.
Innovation Solution
The method employs a tree-based space-partitioning decomposition of the model space to dynamically adjust subdomain boundaries and redistribute computational loads among computer nodes, ensuring balanced load distribution and adaptive rebalancing during simulations, even with changes in available resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If uniform domain decomposition is used with regular gridded schemes, then implementation simplicity is improved, but load balancing performance deteriorates in heterogeneous computing environments
Solution Approach 1:
The patent segments the domain decomposition process into multiple levels using a tree structure, where the root represents the full domain and child nodes represent progressively finer subdomains. This hierarchical segmentation allows the system to balance between simple uniform decomposition at higher levels and customized load-balanced decomposition at lower levels, resolving the contradiction between implementation simplicity and load balancing performance.
Solution Approach 2:
The patent applies local quality by allowing different decomposition strategies at different locations in the domain hierarchy. Leaf nodes in the tree can use customized decomposition methods tailored to local computational characteristics and node capabilities, while parent nodes maintain simpler structures. This enables the system to achieve good load balancing in critical regions without sacrificing overall implementation simplicity.
2Measurement precision
If hybrid modeling kernels with different grid resolutions are used in different regions, then modeling accuracy is improved, but domain decomposition complexity increases
Solution Approach 1:
The patent implements nested doll by embedding finer grid regions within coarser grid regions through the tree structure. Child nodes represent subdomains with higher resolution kernels that are nested within parent node domains using lower resolution kernels. This nesting allows the system to achieve high modeling accuracy in target volumes while maintaining a structured decomposition framework that manages complexity through hierarchical organization.
Solution Approach 2:
The patent adds a hierarchical dimension to the domain decomposition by introducing tree levels. Instead of attempting to manage all decomposition details in a single flat layer, the system organizes subdomains across multiple hierarchical levels, where each level operates at a different resolution. This dimensional transformation allows the system to handle complex hybrid kernels while maintaining manageable decomposition complexity at each level.
3Loss of energy
If static domain decomposition is used before simulation, then computational overhead is reduced, but adaptability to changing resource availability deteriorates
Solution Approach 1:
The patent introduces dynamics by making the domain decomposition adaptable to changing conditions through the tree structure. The hierarchical decomposition allows the system to dynamically adjust subdomain assignments and boundaries in response to changing resource availability, node failures, or load conditions during simulation execution, while maintaining the structural framework that limits computational overhead.
Solution Approach 2:
The patent applies preliminary action by pre-computing and storing decomposition information in the tree structure before simulation begins. The hierarchical decomposition and initial load balancing are performed in advance, creating a ready-to-use framework that can quickly adapt to changes without requiring extensive re-computation during execution. This preliminary preparation reduces operational overhead while maintaining adaptability.
Data Source
AI summary
Methods and apparatuses for domain decomposition in computer simulations using an m-dimensional space-partitioning tree. The domain decomposition may be used in load balancing. Each subdomain boundary is adjusted according to its assigned computer node capability such that its load matches its capability. The subdomain simulation load may be acquired from predictive estimates or from actual measurement during the simulation execution. The load balancing domain decomposition may be done before the simulation starts or during the simulation.