Study of Fine-Grained, Irregular Parallel Applications on a Many-Core Processor

Thumbnail Image


Publication or External Link





This dissertation demonstrates the possibility of obtaining strong speedups for a variety of parallel applications versus the best serial and parallel implementations on commodity platforms. These results were obtained using the PRAM-inspired Explicit Multi-Threading (XMT) many-core computing platform, which is designed to efficiently support execution of both serial and parallel code and switching between the two.

Biconnectivity: For finding the biconnected components of a graph, we demonstrate speedups of 9x to 33x on XMT relative to the best serial algorithm using a relatively modest silicon budget. Further evidence suggests that speedups of 21x to 48x are possible. For graph connectivity, we demonstrate that XMT outperforms two contemporary NVIDIA GPUs of similar or greater silicon area. Prior studies of parallel biconnectivity algorithms achieved at most a 4x speedup, but we could not find biconnectivity code for GPUs to compare biconnectivity against them.

Triconnectivity: We present a parallel solution to the problem of determining the triconnected components of an undirected graph. We obtain significant speedups on XMT over the only published optimal (linear-time) serial implementation of a triconnected components algorithm running on a modern CPU. To our knowledge, no other parallel implementation of a triconnected components algorithm has been published for any platform.

Burrows-Wheeler compression: We present novel work-optimal parallel algorithms for Burrows-Wheeler compression and decompression of strings over a constant alphabet and their empirical evaluation. To validate these theoretical algorithms, we implement them on XMT and show speedups of up to 25x for compression, and 13x for decompression, versus bzip2, the de facto standard implementation of Burrows-Wheeler compression.

Fast Fourier transform (FFT): Using FFT as an example, we examine the impact that adoption of some enabling technologies, including silicon photonics, would have on the performance of a many-core architecture. The results show that a single-chip many-core processor could potentially outperform a large high-performance computing cluster.

Boosted decision trees: This chapter focuses on the hybrid memory architecture of the XMT computer platform, a key part of which is a flexible all-to-all interconnection network that connects processors to shared memory modules. First, to understand some recent advances in GPU memory architecture and how they relate to this hybrid memory architecture, we use microbenchmarks including list ranking. Then, we contrast the scalability of applications with that of routines. In particular, regardless of the scalability needs of full applications, some routines may involve smaller problem sizes, and in particular smaller levels of parallelism, perhaps even serial. To see how a hybrid memory architecture can benefit such applications, we simulate a computer with such an architecture and demonstrate the potential for a speedup of 3.3X over NVIDIA's most powerful GPU to date for XGBoost, an implementation of boosted decision trees, a timely machine learning approach.

Boolean satisfiability (SAT): SAT is an important performance-hungry problem with applications in many problem domains. However, most work on parallelizing SAT solvers has focused on coarse-grained, mostly embarrassing parallelism. Here, we study fine-grained parallelism that can speed up existing sequential SAT solvers. We show the potential for speedups of up to 382X across a variety of problem instances. We hope that these results will stimulate future research.