Exploring the Role of Grid-Enabled OpenMP in Adaptive Mesh Calculations

点击翻译按钮获取中文摘要

Advances in computer and network technologies will soon make it possible to harness the collective computing power of distributed supercomputer systems as a Computational Power Grid (or Grid for short). Grid computing will entail coping with unusually high latencies and low bandwidths. To use such a Grid effectively as a platform for scientific computing will require new algorithms and programming methodologies along with compiler and run-time support to simplify the program development process.The scope of this problem is so broad that we believe that effective technologies can best be developed through focusing on concrete application domains. This proposal will perform an exploratory study to explore the development of compiler and runtime technology as well as algorithmic techniques for supporting largescale, parallel, structured adaptive mesh refinement (AMR) calculations in Computational Fluid Dynamics (CID) on the Grid. The project will be an interdisciplinary study consisting of applications scientists from NCSA and computer scientists from Rice and University of Houston.This proposal, therefore, will carry out a study to see whether such highly complex AMR applications can be efficiently run on the Grid. The course of this work will be exploring several different algorithmic strategies for supporting AMR on Grids. This will also entail developing innovative compiler technology and language extensions to the OpenMP standard that allow it to provide Grid-based support for such AMR applications. Lastly, resources (nodes, network connections) in the Grid computational environment are often contended. Therefore a Grid-enabled application has to be able to speculatively adapt to changing Grid resources. The proposal will build up tools that monitor the application's performance on the Grid as well as the available Grid resources. Information from these tools will be made available to the AMR application through standard interfaces in our OpenMP API so that the AMR application can speculatively adapt to changing Grid resources.