SBIR Phase I: Enhancing the Performance of Scientific Applications Through Intelligent Advice

小型企业创新研究(SBIR)第一阶段项目的更广泛影响/商业潜力将是使新的硬件平台能够更快地部署，并使有用的软件产品更容易在各种硬件上提供，以满足其用户的需求。此外，科学和工程研究人员可以将其应用程序在高性能计算或数据中心内的机器上的运行时间缩短数小时或数天，从而能够在这些中心内完成更科学的建模和模拟，同时允许每位科学家更多地专注于他/她的研究，这反过来可能会更快地造福社会。例如，在国防工业中，许多承包公司开发蒙特卡洛模拟软件来预测灾难的结果。增加这些模拟的吞吐量将允许分析和测试更多的模拟结果，从而更好地预测和处理灾难性情况。这个小型企业创新研究第一阶段项目的独特之处在于，它自动连接现有的性能建模和预测工具，以及模式驱动的编译器优化。将开发新的运行时监控和建模技术，以自动将现有性能分析工具发现的性能瓶颈与源代码优化的潜在机会相对应。这样的机会将再次用于指导模式驱动的编译器优化，特别是在GPGPU和多核/多核CPU上增强有限元方法的性能。深度学习神经网络将用于自动化计算的运行时行为分类。新的模式驱动的编译器优化专业化将使通用编译器技术能够知晓库抽象(例如，数据结构和算法抽象)的更高级别语义，并允许对它们进行集体定制和协调，以获得最佳性能。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase I project will be to allow new hardware platforms to be more quickly deployed and useful software products to be more readily made available on a wide variety of hardware to satisfy the needs of their users. Additionally, researchers in science and engineering can reduce the runtime of their applications by hours or days on the machines inside high performance computing or data centers, enabling more scientific modeling and simulations to be completed within these centers, while allowing each individual scientist to concentrate more on his/her research, which in turn may more quickly benefit society. For example, in the defense industry, many contracting companies develop Monte Carlo simulation software to predict outcomes from disasters. Increasing the throughput of these simulations would allow for more simulation results to be analyzed and tested, thereby resulting in better predictions and handling of disastrous situations. This Small Business Innovation Research Phase I project is unique in its automated connection between existing performance modeling and prediction tools, and pattern-driven compiler optimization. Novel runtime monitoring and modeling techniques will be developed to automatically map performance bottlenecks discovered by existing performance analysis tools to potential opportunities of source code optimizations. Such opportunities again will be used to guide pattern-driven compiler optimizations, particularly to enhance the performance of finite element methods on both GPGPUs and multi-core/many-core CPUs. Deep learning neural networks will be used to automate the runtime behavior classification of computations. The new pattern-driven specialization of compiler optimizations will enable general-purpose compiler techniques to be aware of the higher-level semantics of library abstractions (e.g., data structures and algorithm abstractions) and allow them to be collectively customized and coordinated to attain the best performance.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.