Semantic Relaxation of Concurrent Data Structures: Efficient and Elastic Designs

Licentiatavhandling, 2025

The growing availability of hardware parallelism continues to challenge developers to design programs that efficiently utilize it. Many widely used concurrent data structures — such as FIFO queues and priority queues — enforce strict ordering semantics that inevitably create memory contention, limiting scalability. Semantic relaxation has emerged as a powerful technique for increasing parallelism at the expense of a controlled weakening of the ordering semantics. However, many open questions remain in the field of relaxed data structures, both in the terms of efficient, flexible designs, and in their practical applicability.

This thesis advances the theory and practice of relaxed concurrent data structures. First, we revisit the classic balanced allocations paradigm in the context of queues, introducing the d-CBO relaxed FIFO queue. The d-CBO queue utilizes the classical d-choice in a new way to evenly balance operation counts across sub-queues. Our analysis shows provably low, stable relaxation errors, and experiments demonstrate a better relaxation-performance trade-off than previous relaxed FIFO designs.

Second, we explore the applicability of relaxation in graph analytics. Using a relaxed priority queue in a parallel Single-Source Shortest Path (SSSP) implementation, we achieve state-of-the-art performance on sparse graphs and remain competitive across other graph types. Our findings show that relaxed designs can be used within parallel algorithms to outperform state-of-the-art without extensive parameter tuning or problem-specific tailoring.

Finally, we introduce the concept of elastic relaxation, which enables relaxed implementations to adjust their semantics dynamically during run time. We extend a state-of-the-art framework for relaxed data structures to support elastic variants of queues, stacks, deques, and counters, with correctness guarantees and deterministic relaxation bounds. Experiments show that these elastic capabilities incur minimal overhead. When combined with a lightweight controller for relaxation, they demonstrated an improved trade-off between throughput and work-efficiency compared to static designs.