OpenPME: Open Particle Mesh Environment for Systems Biology

点击翻译按钮获取中文摘要

Parallel high-performance computing (HPC) is essential to study scientific models of complex real-world phenomena, to accurately predict their behavior and reaction to perturbations, and analyze large data sets of real-world experiments. Particularly challenging applications of this kind arise from the field of systems biology,where complex models of cells and systems of cells are simulated in order to reverse-engineer biological mechanisms. The process of modeling and simulation is driven by high-resolution 3D microscopy images, which also need to be processed in real time. Modeling, simulation, and data analysis hence critically depend on effective use of HPC platforms. Developing robust and versatile numerical methods that run efficiently on parallel HPC hardware, however, is a challenging, error-prone, and time-consuming task. Developers need to understand the problem domain at hand, the mathematical models, as well as the impact of algorithmic specifications on the underlying system architecture and performance. Rarely, all of these intricacies can be dealt with in a given project, hence defining a knowledge gap between the application and the HPC system.In this project, we narrow the gap between a high-level problem description and the final HPC code by introducing transparent layers of abstraction, consisting of: a domain-specific language, a compiler, an adaptive runtime system, and a development environment. We focus on generic particle-mesh methods as the numerical simulation framework, because they can simulate both discrete and continuous models and process images, hence covering the entire workflow of systems biology. We will create the first semantic description of particle-mesh algorithms and derive a formal domain-specific language from it. Around this language, we will design a tool-flow and integrated development environment that provides instant user feedback and support for interactive debugging.Our runtime system will be based on the OpenFPM library for particle-mesh methods being currently developed. Leveraging domain knowledge and runtime monitoring, we will develop static and dynamic optimizations at the language level, as well as at runtime. To asses our language and optimizations, three driver applications from computational biology that pose outstanding problems will be implemented and evaluated in terms of runtime performance, scalability, and developers productivity.