Framework for mapping and performance tuning of a weather prediction application on high performance heterogeneous computing systems

• Grant number: 17/04291-0
• Support Opportunities: Research Grants - Visiting Researcher Grant - International
• Start date: June 01, 2017
• End date: June 09, 2017
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computer Systems
• Principal Investigator: Eduardo Marques
• Grantee: Eduardo Marques
• Visiting researcher: Pedro Nuno Cruz Diniz
• Visiting researcher institution: University of Southern California (USC) , United States
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• City of the host institution: São Carlos

Abstract

Emerging architectures are both multi-core and heterogeneous offering power and computational efficiency orders of magnitude beyond traditional high-performance systems. Yet, this heterogeneity exacerbates the complexity of the development of applications codes. To address these issues, we will explore the use of a Domain-Specific Language (DSL), for high-level application composition, compilation and synthesis for hybrid systems comprising CPUs GPUs and FPGAs. This DSL will allow for the design-space exploration of applications with alternative implementations of the same code functionality when targeting distinct computing devices and having distinct tuning parameters. As a case study, we will focus on an application of great significance, the Brazilian Regional Atmospheric Modeling System (BRAMS) which is the operational limited area model providing numerical weather prediction to Brazil. In the exploration of the mapping and tuning of the BRAMS application, we will also leverage existing third-party tools in the 'back-end' of the developed code generation and design exploration tools for hardware synthesis and compilation to minimize risk and technology dependence. The use of a DSL for tuning applications such as the BRAMS, will substantially reduce the effort and time spend tuning them to existing architectures. More importantly, when the target systems are upgraded, the same code tuning strategies can be reused and subsequently refined thus promoting both programmer productivity and program portability. (AU)