UC Irvine

General N-body problems are a set of problems in which an update to a single element in the system depends on every other element. N-body problems are ubiquitous, with applications in various domains ranging from scientific computing simulations in molecular dynamics, astrophysics, acoustics, and fluid dynamics all the way to computer vision, data mining and machine learning problems. Different N-body algorithms have been designed and implemented in these various fields. However, there is a big gap between the algorithm one designs on paper and the code that runs efficiently on a parallel system. It is time-consuming to write fast, parallel, and scalable code for these problems. On the other hand, the sheer scale and growth of modern scientific datasets necessitate exploiting the power of both parallel and approximation algorithms where there is a potential to trade-off accuracy for performance. The main problem that we are tackling in this thesis is how to automatically generate asymptotically optimal N-body algorithms from the high-level specification of the problem. We combine the body of work in performance optimizations, compilers and the domain of N-body problems to build a unified system where domain scientists can write programs at the high level while attaining performance of code written by an expert at the low level.

In order to generate a high-performance, scalable code for this group of problems, we take the following steps in this thesis; first, we propose a unified algorithmic framework named PASCAL in order to address the challenge of designing a general algorithmic template to represent the class of N-body problems. PASCAL utilizes space-partitioning trees and user-controlled pruning/approximations to reduce the asymptotic runtime complexity from linear to logarithmic in the number of data points. In PASCAL, we design an algorithm that automatically generates conditions for pruning or approximation of an N-body problem considering the problem's definition. In order to evaluate PASCAL, we developed tree-based algorithms for six well-known problems: k-nearest neighbors, range search, minimum spanning tree, kernel density estimation, expectation maximization, and Hausdorff distance. We show that applying domain-specific optimizations and parallelization to the algorithms written in PASCAL achieves 10x to 230x speedup compared to state-of-the-art libraries on a dual-socket Intel Xeon processor with 16 cores on real-world datasets. Second, we extend the PASCAL framework to build PASCAL-X that adds support for NUMA-aware parallelization. PASCAL-X also presents insights on the influence of tuning parameters. Tuning parameters such as leaf size (influences the shape of the tree) and cut-off level (controls the granularity of tasks) of the space-partitioning trees result in performance improvement of up to 4.6x.

A key goal is to generate scalable and high-performance code automatically without sacrificing productivity. That implies minimizing the effort the users have to put in to generate the desired high-performance code. Another critical factor is the adaptivity, which indicates the amount of effort that is required to extend the high-performance code generation to new N-body problems. Finally, we consider these factors and develop a domain-specific language and code generator named Portal, which is built on top of PASCAL-X. Portal's language design is inspired by the mathematical representation of N-body problems, resulting in an intuitive language for rapid implementation of a variety of problems. Portal's back-end is designed and implemented to generate optimized, parallel, and scalable implementations for multi-core systems. We demonstrate that the performance achieved by using Portal is comparable to that of expert hand-optimized code while providing productivity for domain scientists. For instance, using Portal for the k-nearest neighbors problem gains performance that is similar to the hand-optimized code, while reducing the lines of code by 68x. To the best of our knowledge, there are no known libraries or frameworks that implement parallel asymptotically optimal algorithms for the class of general N-body problems and this thesis primarily aims to fill this gap. Finally, we present a case study of Portal for the real-world problem of face clustering. In this case study, we show that Portal not only provides a fast solution for the face clustering problem with similar accuracy as the state-of-the-art algorithm, but also it provides productivity by implementing the face clustering algorithm in only 14 lines of Portal code.