SHF: Medium: Collaborative Research: An Inspector/Executor Compilation Framework for Irregular Applications

点击翻译按钮获取中文摘要

Computational science and engineering provides inexpensive exploration of physical phenomena and design spaces and helps direct experimentation and advise theory. Irregular applications such as molecular dynamics simulations, n-body simulations, finite element analysis, and big graph analytics constitute a critical and significant portion of scientific computing applications. An irregular application is characterized by having indirect memory accesses that cannot be determined when the application is being compiled, therefore severely limiting the applicability of the large body of work on parallelizing compiler technology. Consequently, irregular applications, which are so important in pushing forward the frontiers of science, place a very large burden on computational and domain scientists in developing high-performance implementations for the ever-changing landscape of parallel architectures. The intellectual merit of this project is to develop a compiler and runtime framework for irregular applications, particularly well suited for sparse matrix and graph computations that underlie critical problems in computational science and data science. The broader impact is to provide domain scientists a powerful tool for optimizing and porting performance-critical, irregular computations to current and future multi-core processors and many-core accelerators. The PIs will also continue efforts in outreach and diversity to increase the participation in STEM careers, particularly among women and underrepresented minorities.The approach in this project is to extend the well-established inspector/executor paradigm where the computational dependence structure (based on the memory access pattern) is determined at runtime, and runtime information is passed to a compile-time generated executor. Specifically, an inspector can examine the memory access patterns early in the computation at runtime, and an executor leverages this information to perform data and computation reordering and scheduling to affect memory hierarchy and parallelism optimizations. The project is developing a compiler and runtime framework with new abstractions for expressing and manipulating inspectors; these inspectors may then be integrated nearly seamlessly with each other and with existing compiler optimizations (e.g., loop tiling) to optimize executors. The project is also extending prior work that supports non-affine input code and mixes compile-time and runtime optimization. The resulting system increases the productivity of expert programmers in achieving both high performance and portability on a wide variety of irregular applications.