The Benefits of Adaptive Partitioning for Parallel AMR Applications

Technical Report · Tue Jul 01 04:00:00 EDT 2008

DOI:https://doi.org/10.2172/1324617· OSTI ID:1324617

Steensland, Johan ^[1]

---

1. Sandia National Lab. (SNL-CA), Livermore, CA (United States). Advanced Software Research and Development

Parallel adaptive mesh refinement methods potentially lead to realistic modeling of complex three-dimensional physical phenomena. However, the dynamics inherent in these methods present significant challenges in data partitioning and load balancing. Significant human resources, including time, effort, experience, and knowledge, are required for determining the optimal partitioning technique for each new simulation. In reality, scientists resort to using the on-board partitioner of the computational framework, or to using the partitioning industry standard, ParMetis. Adaptive partitioning refers to repeatedly selecting, configuring and invoking the optimal partitioning technique at run-time, based on the current state of the computer and application. In theory, adaptive partitioning automatically delivers superior performance and eliminates the need for repeatedly spending valuable human resources for determining the optimal static partitioning technique. In practice, however, enabling frameworks are non-existent due to the inherent significant inter-disciplinary research challenges. This paper presents a study of a simple implementation of adaptive partitioning and discusses implied potential benefits from the perspective of common groups of users within computational science. The study is based on a large set of data derived from experiments including six real-life, multi-time-step adaptive applications from various scientific domains, five complementing and fundamentally different partitioning techniques, a large set of parameters corresponding to a wide spectrum of computing environments, and a flexible cost function that considers the relative impact of multiple partitioning metrics and diverse partitioning objectives. The results show that even a simple implementation of adaptive partitioning can automatically generate results statistically equivalent to the best static partitioning. Thus, it is possible to effectively eliminate the problem of determining the best partitioning technique for new simulations. Moreover. the results show that adaptive partitioning can provide a performance gain of about 10 percent on average as compared to routinely using the industry-standard, ParMetis.

• Research Organization: Sandia National Lab. (SNL-CA), Livermore, CA (United States)
• Sponsoring Organization: USDOE National Nuclear Security Administration (NNSA). Advanced Simulation and Computing (ASC) Program
• DOE Contract Number: AC04-94AL85000
• OSTI ID: 1324617
• Report Number(s): SAND2008--4073; 278442
• Country of Publication: United States
• Language: English

Related Subjects

97 MATHEMATICS AND COMPUTING