- Datum:Startar 11 maj 2023, 10:00Slutar 11 maj 2023, 12:00
Efficient concurrent data structure access parallelism techniques for increasing scalability
Multi-core processors have revolutionised the way data structures are designed by bringing parallelism to mainstream computing. Key to exploiting hardware parallelism available in multi-core processors are concurrent data structures. However, some concurrent data structure abstractions are inherently sequential and incapable of harnessing the parallelism performance of multi-core processors. Designing and implementing concurrent data structures to harness hardware parallelism is challenging due to the requirement of correctness, efficiency and practicability under various application constraints. In this thesis, our research contribution is towards improving concurrent data structure access parallelism to increase data structure performance. We propose new design frameworks that improve access parallelism of already existing concurrent data structure designs. Also, we propose new concurrent data structure designs with significant performance improvements. To give an insight into the interplay between hardware and concurrent data structure access parallelism, we give a detailed analysis and model the performance scalability with varying parallelism.
In the first part of the thesis, we focus on data structure semantic relaxation. By relaxing the semantics of a data structure, a bigger design space, that allows weaker synchronization and more useful parallelism, is unveiled. Investigating new data structure designs, capable of trading semantics for achieving better performance in a monotonic way, is a major challenge in the area. We algorithmically address this challenge in this part of the thesis. We present an efficient, lock-free, concurrent data structure design framework for out-of-order semantic relaxation. We introduce a new two-dimensional algorithmic design, that uses multiple instances of a given data structure to improve access parallelism.
In the second part of the thesis, we propose an efficient priority queue that improves access parallelism by reducing the number of synchronization points for each operation. Priority queues are fundamental abstract data types, often used to manage limited resources in parallel systems. Typical proposed parallel priority queue implementations are based on heaps or skip lists. In recent literature, skip lists have been shown to be the most efficient design choice for implementing priority queues. Though numerous intricate implementations of skip list based queues have been proposed in the literature, their performance is constrained by the high number of global atomic updates per operation and the high memory consumption, which are proportional to the number of sub-lists in the queue. In this part of the thesis, we propose an alternative approach for designing lock-free linearizable priority queues, that significantly improve memory efficiency and throughput performance, by reducing the number of global atomic updates and memory consumption as compared to skip-list based queues. To achieve this, our new design combines two structures; a search tree and a linked list, forming what we call a Tree Search List Queue (TSLQueue).
Subsequently, we analyse and introduce a model for lock-free concurrent data structure access parallelism. The major impediment to scaling concurrent data structures is memory contention when accessing shared data structure access points, leading to thread serialisation, and hindering parallelism. Aiming to address this challenge, a significant amount of work in the literature has proposed multi-access techniques that improve concurrent data structure parallelism. However, there is little work on analysing and modelling the execution behaviour of concurrent multi-access data structures especially in a shared memory setting. In this part of the thesis, we analyse and model the general execution behaviour of concurrent multi-access data structures in the shared memory setting. We study and analyse the behaviour of the two popular random access patterns: shared (Remote) and exclusive (Local) access, and the behaviour of the two most commonly used atomic primitives for designing lock-free data structures: Compare and Swap, and, Fetch and Add.