Greedy Sharing: Load Balancing on Weakly Consistent Memory

An efficient online scheduler is crucial for balancing irregular parallel computations in a multiprocessor system. Over the last two decades, variants of the work-stealing scheduler have emerged as a popular choice for hardware shared-memory systems. The state-of-the-art work-stealing algorithms can guarantee near-optimal asymptotic complexity by relying on simple yet powerful techniques to balance total load among processors. Implementations of work stealing algorithms, however, continue to rely on synchronization operations, such as atomic read-write operations (e.g., compare and swap), to guarantee correctness of concurrent accesses to shared task pools. Furthermore, since work-stealing algorithms are traditionally designed by assuming a sequentially-consistent memory model, their implementations use additional memory fences on modern multiprocessor machines. Memory fences and atomic-read-write operations are known to be expensive in general, especially because they can require exclusive access to shared resources, such as the memory. In this paper, we present the greedy-sharing algorithm for load balancing on weakly-consistent hardware shared-memory architectures, such as modern multicore computers. Greedy sharing combines ideas from work-sharing and work-stealing algorithms to eliminate all synchronization operations. As in work sharing, busy processors perform load balancing by sharing their work; as in work-stealing, tasks migrate only from busy to idle processors. In greedy sharing, data races can occur when multiple processors target the same idle processor. To recover safely from such data races, we design a protocol for task sharing that makes use of particular time-stamping technique, which is attributed to Lamport. We present a specification of the algorithm, prove its correctness on the X86-TSO weak memory model, and prove an upper bound for its execution time. We have implemented the algorithm as part of our C++ library for parallel programming. We present experiments to show that the algorithm is practical.