Improving basic thread operations with batches of threads

Multi-core architectures provide the means to efficiently handle more fine-grained and larger numbers of parallel tasks. However, software still does not take advantage of these new possibilities, retaining the high cost associated with managing large numbers of threads. Batches of Threads have been introduced to reduce this cost and allow applications to express their inherent parallelism in a more fine-grained manner. In this paper, their use is extended, in order to improve two significant aspects of threading run-time systems. Firstly, to schedule large numbers of threads to processors. Secondly, to recycle data structures of threads that have finished execution. Both improvements can be implemented internally in threading run-time systems and thus are transparent to the programmer. The experimental evaluation demonstrates that basic thread operations improve significantly.