SRC06. Toward Exascale Computing: An Automatic Hardware-Software Co-Design Framework for Aspen
Student: Mariam Umar (Virginia Polytechnic Institute and State University)
Supervisor: Kirk W. Cameron (Virginia Polytechnic Institute and State University)
Abstract: Hardware-software co-design is becoming a constraint as we approach the exascale era, particularly because of increases in complexity and scale. Existing solutions to maximize performance within power constraints, like memory throttling and DVFS, are becoming less efficient with generational increases in type and density of cores and accelerators. This problem is exacerbated when implementing application specific hardware. We, therefore propose automated generation of performance models for hardware and software to simplify the co-design process. Our hardware-software co-design framework is developed for Aspen: a domain specific language designed for testing and analyzing exascale systems. The framework uses application and hardware information, extracts their characteristics, generates intermediate representations, verifies them and produces Aspen models. Our approach has numerous applications, e.g., sensitivity analysis, estimating and comparing performance models on current and future architectures. Here we present sensitivity analysis of six applications, and scaling and energy profiling for Titan and future exascale system.
Two-page extended abstract: pdf