Parallel Processor Geophysical Data Inversion Load Balancing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for joint geophysical data inversion on parallel processors face inefficiencies due to load imbalance, memory constraints, and varying computational demands across different data types, leading to suboptimal performance and increased processing times.
Innovation Solution
A method is implemented where a control processor manages operations to reduce load imbalance, with dedicated processor groups for updating model parameters and simulating forward modeling data, allowing for iterative optimization that compares simulated and measured data to minimize misfit, and dynamically assigns tasks based on spatial parallelization, source, or receiver location.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a static decomposition of data among processor groups is used, then the system structure is simple, but load imbalance occurs causing significant variations in simulation time
Solution Approach 1:
The patent implements dynamic load balancing where the controller redistributes simulation tasks among processor groups based on their current workload and performance characteristics. This dynamic adjustment allows the system to adapt to varying computational demands during iterative inversion processes, preventing load imbalance without requiring a fundamentally complex system architecture.
Solution Approach 2:
The system performs preliminary characterization of processor group performance before the main inversion process. This pre-assessment allows the controller to make informed decisions about task distribution, optimizing simulation time efficiency while maintaining relatively simple system structures.
2Ease of operation
If the resistivity model is copied to all processor groups, then forward simulations can be performed independently, but memory availability decreases
Solution Approach 1:
The patent divides the resistivity model into segments that are distributed across processor groups rather than copying the complete model to all processors. This segmentation allows each processor group to work with only the portion of the model it needs, maintaining independent simulation capability while significantly reducing total memory requirements.
Solution Approach 2:
The controller acts as an intermediary that manages data exchange between processor groups. Instead of requiring all processors to have complete model copies, the controller coordinates the distribution and retrieval of model segments, enabling independent simulations with reduced memory footprints.
3Quantity of substance
If more processors are assigned to each group to compensate for memory limitations, then the number of concurrent simulations decreases, but individual simulation capacity increases
Solution Approach 1:
By segmenting the resistivity model and distributing it across multiple processor groups, the system enables more groups to participate in concurrent simulations. Each group handles a portion of the model, allowing the total number of concurrent simulations to increase while maintaining adequate memory capacity per group.
Solution Approach 2:
The patent introduces a new dimension of parallelism by distributing the model spatially across processor groups rather than increasing computational power within each group. This dimensional shift from intra-group to inter-group parallelism allows more simulations to run concurrently while keeping individual group memory requirements manageable.
4Ease of operation
If frequency is used as a proxy for simulation time for load balancing, then task assignment is simplified, but accuracy of load balancing decreases
Solution Approach 1:
The system implements feedback mechanisms where the controller monitors actual simulation performance and uses this information to adjust task assignments. This feedback loop maintains load balancing accuracy without significantly complicating the task assignment process, as the controller automatically adapts based on observed performance patterns.
Solution Approach 2:
The system performs preliminary performance characterization of processor groups to establish more accurate task assignment strategies. This pre-assessment reduces reliance on simple frequency proxies and enables more accurate load balancing from the outset, minimizing time losses.
Data Source
AI summary
A method for efficient use of a computing system of parallel processors to perform inversion of geophysical data, or joint inversion of two or more data types. The method includes assigning at least one control processor to control sequence of operations and reduce load imbalance, assigning a group of one or more processors dedicated to updating one or more model parameters, and assigning another group of one or more processors dedicated to forward modeling simulated data.