Embedded Genetic Allocator.

Abstract : This final report documents the work accomplished under the Embedded Genetic Allocator Program, a two year effort funded by DARPA/ITO. The major accomplishment of the program was to develop a new approach to the problem of automatically optimizing the use of memory and processor resources in high performance computing systems consisting of heterogeneous processor nodes connected on a high-speed interconnection fabric. This is frequently known as the mapping problem. The Embedded Genetic Allocation technology developed under this program can provide an automated mapping tool for the design and re-hosting of these systems. The approach developed is automatic and broadly applicable to a wide variety of system architectures. It consists of a hybrid genetic algorithm optimizer (the Embedded Genetic Allocator or EGA), coupled with a software performance monitoring system (various ones can be used). The results presented in this report demonstrate that the EGA can be used to optimize the allocation mappings real-world software, and that the resulting optimizations can rival or even improve upon those generated manually by a skilled programmer.