Implementation and Scalability of a Pure Java Parallel Framework with Application to Hyperbolic Conservation Laws (Preprint)
Abstract
We introduce a pure Java parallel framework for Single Process, Multiple Dataset (SPMD) applications, intended for time-accurate solutions of hyperbolic conservation laws. The software architecture is based upon an extension of the client-server paradigm, utilizing a tree database abstraction and allowing for multi-tiered network configurations. The framework is designed to be hardware independent, with the ability to handle both shared-memory and distributed-memory hardware alike, allowing execution over heterogeneous networks. Task division is determined through permanent domain decomposition, in which Java threads are created for each computation domain and are then distributed over the available servers. Java Remote Method Invocation (RMI) is used for network-based communication of critical I/O as well as thread communication and cooperation between different Java Virtual Machines (JVM). Parallel efficiency and scalability of the framework for both shared-memory and distributed-memory hardware are evaluated for standardized benchmark problem computed with the Euler equations of gas dynamics. Results show efficient use of multiple process or resources on shared-memory systems with minimal thread overheads and near linear scalability on distributed networks with up to 50 server nodes (100 processors).
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 04, 2008
- Accession Number
- ADA486114
Entities
People
- Jean Luc Cambier
- Michael Kapper
Organizations
- Air Force Research Laboratory