Date

2015

Document Type

Dissertation

Degree

Doctor of Philosophy

Department

Computer Science

First Adviser

Spear, Michael

Other advisers/committee members

Spear, Michael; Korth, Hank; Tan, Gang; Luchangco, Victor

Abstract

Concurrent data structures lie at the heart of modern parallel programs. The design and implementation of concurrent data structures can be challenging due to the demand for good performance (low latency and high scalability) and strong progress guarantees. In this dissertation, we enrich the knowledge of concurrent data structure design by proposing new implementations, as well as general techniques to improve the performance of existing ones.The first part of the dissertation present an unordered linked list implementation that supports nonblocking insert, remove, and lookup operations. The algorithm is based on a novel ``enlist'' technique that greatly simplifies the task of achieving wait-freedom. The value of our technique is also demonstrated in the creation of other wait-free data structures such as stacks and hash tables.The second data structure presented is a nonblocking hash table implementation which solves a long-standing design challenge by permitting the hash table to dynamically adjust its size in a nonblocking manner. Additionally, our hash table offers strong theoretical properties such as supporting unbounded memory. In our algorithm, we introduce a new ``freezable set'' abstraction which allows us to achieve atomic migration of keys during a resize. The freezable set abstraction also enables highly efficient implementations which maximally exploit the processor cache locality. In experiments, we found our lock-free hash table performs consistently better than state-of-the-art implementations, such as the split-ordered list.The third data structure we present is a concurrent priority queue called the ``mound''. Our implementations include nonblocking and lock-based variants. The mound employs randomization to reduce contention on concurrent insert operations, and decomposes a remove operation into smaller atomic operations so that multiple remove operations can execute in parallel within a pipeline. In experiments, we show that the mound can provide excellent latency at low thread counts.Lastly, we discuss how hardware transactional memory (HTM) can be used to accelerate existing nonblocking concurrent data structure implementations. We propose optimization techniques that can significantly improve the performance (1.5x to 3x speedups) of a variety of important concurrent data structures, such as binary search trees and hash tables. The optimizations also preserve the strong progress guarantees of the original implementations.

Share

COinS