Efficient and Practical Non-Blocking Data Structures
Doctoral thesis, 2004
This thesis deals with how to design and implement efficient, practical and reliable concurrent data structures. The design method using mutual exclusion incurs serious drawbacks, whereas the alternative non-blocking techniques avoid those problems and also admit improved parallelism. However, designing non-blocking algorithms is a very complex task, and a majority of the algorithms in the literature are either inefficient, impractical or both.
We have studied how information available in real-time systems can improve and simplify non-blocking algorithms. We have designed new methods for recycling of buffers as well as time-stamps, and have applied them on known non-blocking algorithms for registers, snapshots and priority queues.
We have designed, to the best of our knowledge, the first practical lock-free algorithm of a skip list data structure. Using our skip list construction we have designed a lock-free algorithm of the priority queue abstract data type, as well as a lock-free algorithm of the dictionary abstract data type.
We have designed, to the best of our knowledge, the first practical lock-free algorithm of a doubly linked list data structure. The algorithm supports well-defined traversals in both directions including deleted nodes. Using our doubly linked list construction we have designed a lock-free algorithm of the deque abstract data type. For the lock-free algorithms presented in this thesis, we have given correctness proofs of the strong consistency property called linearizability and the non-blocking properties.
We have made implementations for actual systems of the algorithms presented in this thesis and a representative set of related non-blocking as well as lock based algorithms in the literature. We have built a framework that combines the implementations in the form of a software library that offers a unified and efficient interface in combination with a portable design.
We have performed empirical performance studies of the data structures presented in this thesis in comparison with related alternative solutions. The experiments performed on an extensive set of multi-processor systems show significant improvements for non-blocking alternatives in general, and for the implementations presented in this thesis in particular.
shared data structure
abstract data type
doubly linked list