TL;DR
This paper introduces a parallel processing approach to significantly speed up probabilistic fracture mechanics analyses of gas turbine disks. By distributing the analysis of different disk regions across a network of computers, they achieve substantial reductions in computation time, making it more feasible to evaluate complex rotor designs and assess their reliability.
What Problem Does This Paper Address?
Designing gas turbine disks for aircraft involves ensuring their structural integrity under extreme conditions. A critical aspect of this is assessing the probability of fracture, especially considering potential defects within the disk material. Traditional methods for this probabilistic analysis, which involve complex simulations and Monte Carlo methods, are computationally intensive and time-consuming. This limits the ability of engineers to perform extensive design iterations and explore various scenarios. The paper addresses the need for a more efficient way to conduct these analyses, particularly for complex geometries like gas turbine disks that are susceptible to metallurgical defects.
Methods
The core of the paper's contribution is a spatial domain decomposition approach for parallelizing the existing DARWIN software. DARWIN is a probabilistic damage tolerance analysis tool that integrates stress analysis, fracture mechanics, and probabilistic modeling.
Here's how their parallel method works:
- Zone-Based Discretization: The gas turbine disk is divided into numerous "zones." Each zone is a region of material assumed to have a uniform stress state and consistent properties for inspection and defect analysis.
- Parallel Analysis: Instead of running the entire analysis on a single machine, each zone (or a group of zones) is assigned to a separate computer in a network. This is achieved by creating individual input files for each "worker" computer.
- Distributed Computing: The UTSA parallel processing network, managed by the Condor job scheduler, distributes these zone-specific tasks to available computers. Condor handles job scheduling, sending the necessary DARWIN executable, input files, and finite element data to worker machines and retrieving the results.
- Master-Worker Scheme: A master computer orchestrates the process, sending out jobs and collecting results. Condor's master-worker scheme ensures that if a worker machine becomes unavailable, the job is reassigned.
- Result Aggregation: Once all individual zone analyses are complete, a master input file combines these results to compute the overall probability of fracture for the entire disk.
Datasets
Not discussed in the paper. The paper focuses on the methodology and its application to a simulated gas turbine disk problem.
Key Insights and Ideas
- Granularity Matters: The paper highlights the "granularity problem" – how to best split the analysis into manageable chunks (files) for parallel processing. Too many small files lead to excessive communication overhead, while too few files don't fully utilize the available parallel resources.
- Network Heterogeneity is Manageable: The UTSA network consists of various computers (PCs, SGI Origin), and Condor's dynamic load balancing effectively handles this heterogeneity, assigning jobs to available resources.
- Communication Overhead is a Bottleneck: For a large number of parallel runs, communication time can significantly erode the benefits of parallel processing, leading to reduced efficiency. This emphasizes the importance of optimizing the size of individual job files.
- Optimum File Size is Crucial: The study found an optimal number of zones per input file (around 52 zones for their test case) that balances computational load and communication overhead, leading to the highest speedup and efficiency.
Key Results
The paper demonstrates significant performance improvements using their parallel processing approach:
- Speedup: For an extreme test case with 6250 zones, they achieved a speedup of 13.9 when using 121 files (each containing approximately 52 zones) on 22 computers. This means the parallel solution was 13.9 times faster than a serial execution.
- Efficiency: The maximum efficiency achieved was 77%, indicating that the parallel algorithm effectively utilized the available processors.
- Optimum File Size: The analysis suggested that an optimal number of input files is a small multiple (e.g., 3 or 4) of the number of available computers. For their test case, approximately 120 files (around 52 zones per file) provided the best results.
- Communication Impact: With a very large number of files (6250), efficiency dropped to as low as 8% due to communication overhead.
Limitations and Future Work
Limitations:
- Communication Overhead: As noted, communication time becomes a significant bottleneck with a very large number of small files, reducing efficiency.
- Network Availability: The performance of the parallel network is dependent on the availability of lab computers, which are also used for teaching and research. This leads to fluctuations in the number of available computers throughout the day.
Future Work:
- Multi-threading for Single Zones: The authors plan to explore parallelization within a single zone analysis using multi-threading (e.g., with OpenMP) for shared-memory multi-processor systems. This would represent another level of parallelization.
- Parameter Studies and Verification: The parallel network will be used for parameter studies (e.g., investigating load histories, inspection times) and to quickly run verification problems for the DARWIN software.
Summary updated