Scheduling Macro-DataFlow Programs on Task-Parallel Runtime Systems

Though multicore systems are ubiquitous, parallel programming models for these systems are generally not accessible to a wide programmer community. The macro-dataflow model is an attractive stepping stone to implicit parallelism for domain experts who are not the target audience for explicit parallel programming models. We use Intel's Concurrent Collections (CnC) programming model as a concrete exemplar of the macro-dataflow model in this work. CnC is a high level coordination language that can be implemented on top of lower-level task-parallel frameworks. In this thesis, we study an implementation of CnC, based on Habanero-Java as the underlying task-parallel runtime system. A unique feature of CnC, first-class decoupling of data and control dependences, allows us to experiment with schedulers by taking these data and control dependences into account for better scheduling decisions. Our observations led to the proposal and implementation of a new task-parallel synchronization construct for Habanero-Java, namely Data-Driven Futures. We obtained two kinds of experimental results from our implementation. First, we compare the effectiveness of task scheduling policies for CnC programs. Secondly, we show that data-driven futures not only reduce execution time but also shrink memory footprint. In summary, this thesis shows a macro-dataflow programming model can deliver productivity and performance on modern multicore processors