Efficient runtime thread management for the nano-threads programming model

Abstract. The nano-threads programming model was proposed to effectively in-tegrate multiprogramming on shared-memory multiprocessors, with the exploita-tion of fine-grain parallelism from standard applications. A prerequisite for the applicability of the nano-threads programming model is the ability of the runtime environment to manage parallelism at any level of granularity with minimal over-heads. In this paper, we introduce runtime techniques for efficient memory man-agement and user-level scheduling in an experimental runtime system designed to support the nano-threads programming model. We evaluate the exploitation of processor affinity for the management of nano-thread contexts, and the use of hierarchical queues to implement user-level scheduling strategies for applica-tions with inherent multilevel parallelism. The proposed mechanisms attempt to obtain maximum benefits from data locality on cache-coherent NUMA multipro-cessors. Through the use of synthetic benchmarks, we find that our mechanism for memory management in the runtime system reduces overheads by 52 % on average, compared to other known mechanisms. The use of hierarchical queues gives significant performance improvements between 17 % and 40%, compared to scheduling strategies that use local queues