A Lock-Free Work-Stealing Algorithm for Bulk Operations

A team of computer scientists has published a new research paper on arXiv detailing a specialized lock-free work-stealing algorithm designed for bulk operations. The work, led by Raja Sai Nandhan Yadav Kataru, Danial Davarnia, and Ali Jannesari, targets the specific concurrency model of a master-worker framework used to parallelize a mixed-integer programming optimization solver based on decision diagrams. Unlike general-purpose parallel runtimes, this algorithm is built for a restricted model with at most one owner and one concurrent stealer, allowing it to eliminate the overhead of heavy synchronization mechanisms. The design natively supports bulk operations and can grow without bounds, providing a streamlined approach for specialized computational workloads.

Benchmark results demonstrate significant performance advantages over existing general-purpose solutions. The implementation achieves constant-latency push performance, remaining stable even as batch size increases—a stark contrast to queues from C++ Taskflow, whose latencies grow sharply. While pop operations perform comparably, the new algorithm's steal operation maintains nearly flat latency across different steal proportions. The researchers also explored an optimized steal variant that reduces latency by up to 3x in practice. Although a pseudo workload based on large-graph exploration showed all implementations scaling linearly, the authors argue that solver workloads with highly irregular node processing times would further amplify the advantages of their specialized, lightweight design.