Parallel Solver Hierarchy for Fault-Tolerant High-Performance Computing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current fault-resilient methods for high-performance computing systems are inadequate, particularly in exascale computing, where the likelihood of component failures is high, and existing solutions like automatic MPI responses, checkpointing, and replication are either inefficient or not applicable to all types of applications, especially those not based on grids.
Innovation Solution
Implementing a hierarchy of solvers within computing software, where a less accurate solver is used for fault resilience, allowing substitution of values from a lower hierarchy solver in case of failures, enabling the system to maintain accuracy and efficiency even with node failures, and allowing the computation to complete without significant slowdown.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If checkpointing is used for fault tolerance, then reliability is improved, but loss of time increases due to frequent saving and restarting
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing correction values in lookup tables before the main simulation runs. When a fault occurs, these pre-prepared correction values can be immediately applied without restarting computations, thus improving reliability while minimizing time loss.
Solution Approach 2:
The patent implements beforehand cushioning by creating redundant correction data structures and lookup tables in advance. These cushioning resources are prepared beforehand to compensate for potential faults, allowing the system to recover from failures without significant time penalty while maintaining reliability.
2Reliability
If replication of work is used for fault tolerance, then reliability is improved, but device complexity increases due to redundant resources
Solution Approach 1:
The patent uses copying by creating compact correction value tables that store the essential information needed for fault recovery. Instead of replicating entire computational work, only critical correction data is copied and stored, achieving reliability improvement with minimal increase in device complexity.
Solution Approach 2:
The patent applies this principle by using lightweight, easily regenerable correction value tables instead of expensive, complex redundant computational infrastructure. These correction tables are simple data structures that can be quickly recreated if needed, providing fault tolerance without significant device complexity.
3Reliability
If a hierarchy of solvers is implemented for fault tolerance, then reliability is improved, but device complexity increases due to multiple solver levels
Solution Approach 1:
The patent applies segmentation by dividing the computational problem into multiple hierarchy levels with different solver complexities. Each level handles specific types of computations, and the correction value tables store results from lower-level solvers to compensate for faults in higher-level solvers, improving reliability while managing software complexity through structured division.
Solution Approach 2:
The patent implements universality by designing a multi-purpose hierarchy of solvers where each solver level can serve multiple functions. The correction value tables generated by lower-level solvers serve both as computational results and as fault compensation data, allowing the same hierarchical structure to provide both performance optimization and fault tolerance without proportionally increasing complexity.
Data Source
AI summary
A method of improving fault tolerance in a computing system arranged to find a computational solution, the method comprising: computing at least two versions of the solution by using a hierarchy of at least two different solvers in parallel; and if there is a fault during execution of a solver resulting in a missing value, substituting a value from a solver that is lower in the hierarchy to replace the missing value.