Applied Scientific Research — Department of Defense SBIR Phase II: The objective of this proposal is to develop an efficient user-friendly engineering toolki

Description

The objective of this proposal is to develop an efficient user-friendly engineering toolkit consisting of a set of FORTRAN callable routines for desktop solution of variable-coefficient Poisson equations on shared-memory Multi-Core Processors (MCPs). The software design is highly modular, consisting of a flexible user interface linked to a set of hardware-optimized shared-memory Krylov solvers with multigrid and other preconditioners. During Phase I, using a benchmark of interest to the Army, various preconditioner-solver combinations were tested for optimal convergence and solution times, cache-aware data storage formats were examined for scalability, the bottleneck in MCP computing was identified, and the feasibility of a new proposed strategy was demonstrated successfully. For Phase II, the Krylov solvers and preconditioners will be optimized further using a range of information on the underlying MCP hardware. Each software module will be tested rigorously and the integrated system finally validated using applications of interest to the Army. Fast mixed-precision computations, efficient utility of the MCP memory hierarchy and distribution, flexible user interface, and dynamic solver adaptivity characterize the novel contributions of this project to the art of solving variable-coefficient Poisson equations on shared-memory MCPs.