TraceSymm: Trace analysis and Symmetry theory for improved application mapping onto manycores

点击翻译按钮获取中文摘要

The diminishing returns in single-core architectures around 2005 marked an inflection point in computing systems. Since then, multi-cores in all kinds of devices proliferated, causing a boom in models and methodologies focused on how to efficiently program parallel architectures. While industry opted for extensions to established languages (e.g., OpenMP), programming models with formal properties are more popular in academia. In particular, dataflow programming models from the 1970s recently regained interest and are today a matter of a large body of research in the embedded and general purpose domains. A central research problem is that of computing an optimal mapping from a high-level application description to a parallel hardware platform, including the placement of computation to cores, communication to memories and interconnect, and deciding on the execution order.In the meantime, applications have evolved themselves so that they cannot always be described by static models, i.e., for which static analysis suffices to compute an optimal mapping. A common approach to handle this dynamism consists in analyzing executiontraces of an application to better tailor the mapping. With upcoming larger systems and more dynamic applications it is essential to improve trace analysis and trace-based mapping methodologies. In this project we will study, implement and benchmark (i) better trace and hardware representations for scalability, (ii) trace formalisms and analysis algorithms, (iii) symmetry analysis for applications, architectures and mappings, and (iv) novel trace and symmetry-aware mapping algorithms.