Scalable Scheduling for Seismic Inversion on Parallel Cores
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for iterative seismic inversion are computationally expensive and inefficient, especially when scaling to large numbers of cores, due to high communication overhead and lack of reliable scheduling, leading to delays and low hardware utilization in high-performance computing systems.
Innovation Solution
A distributed parallel framework for scalable and reliable scheduling of iterative geophysical inversions, utilizing a pool of independent workers and a central dispatcher to manage task queues, enforce dependencies, and monitor the inversion state, allowing for dynamic resource allocation and failure recovery, while decoupling computation and control mechanisms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If iterative seismic inversion is performed using traditional parallel computing methods, then computational accuracy is improved, but communication overhead increases and system efficiency decreases when scaling to large numbers of cores
Solution Approach 1:
The patent segments the iterative seismic inversion computation into independent worker tasks that can execute autonomously. Each worker processes a portion of the inversion independently, reducing the need for frequent communication between processing units. This segmentation allows the system to maintain computational accuracy while minimizing communication overhead when scaling to large numbers of cores.
2Power
If traditional parallel computing methods are used for iterative seismic inversion, then computational power is increased, but hardware utilization decreases due to delays and scheduling inefficiencies
Solution Approach 1:
The patent implements dynamic scheduling where the central dispatcher continuously monitors worker status and dynamically assigns new tasks as workers become available. This dynamic approach ensures that computational power is fully utilized by keeping workers continuously engaged, eliminating idle time and maximizing hardware utilization while maintaining high computational power.
3Measurement precision
If iterative seismic inversion runs are extended to improve convergence, then model accuracy is improved, but computation time increases and reliability decreases due to potential failures
Solution Approach 1:
The patent implements preliminary checkpointing where the central dispatcher periodically saves the inversion state before extended computation runs. This preliminary action allows the system to recover from failures without losing significant progress, thereby improving reliability while still allowing extended computation times to achieve the necessary model accuracy for convergence.
4Speed
If more cores are added to increase computational speed, then processing speed is improved, but coordination complexity and communication overhead increase
Solution Approach 1:
The patent introduces a central dispatcher as an intermediary between workers and the inversion control system. This mediator manages task assignment, state collection, and coordination for all workers, simplifying the complexity of coordinating hundreds or thousands of cores. The central dispatcher abstracts the coordination complexity, allowing processing speed to scale with the number of cores without proportionally increasing coordination complexity.
Data Source
AI summary
System and method for scalable and reliable scheduling of iterative seismic full wavefield inversion algorithms with alternating parallel and serial stages of computation on massively parallel computing systems. The workers are independent, initiating actions and unaware of each other, and given limited information. This enables application of optimal scheduling, load-balancing, and reliability techniques specific to seismic inversion problems. The central dispatcher specifies the structure of the inversion, including task dependency, and keeps track of progress of work. Management tools enable the user to make performance and reliability improvements during the execution of the seismic inversion.