Extended collectives library for unified parallel C

[Abstract] Current multicore processors mitigate single-core processor problems (e.g., power, memory and instruction-level parallelism walls), but they have raised the programmability wall. In this scenario, the use of a suitable parallel programming model is key to facilitate a paradigm shift from sequential application development while maximizing the productivity of code developers. At this point, the PGAS (Partitioned Global Address Space) paradigm represents a relevant research advance for its application to multicore systems, as its memory model, with a shared memory view while providing private memory for taking advantage of data locality, mimics the memory structure provided by these architectures. Unified Parallel C (UPC), a PGAS-based extension of ANSI C, has been grabbing the attention of developers for the last years. Nevertheless, the focus on improving performance of current UPC compilers/ runtimes has been relegating the goal of providing higher programmability, where the available constructs have not always guaranteed good performance. Therefore, this Thesis focuses on making original contributions to the state of the art of UPC programmability by means of two main tasks: (1) presenting an analytical and empirical study of the features of the language, and (2) providing new functionalities that favor programmability, while not hampering performance. Thus, the main contribution of this Thesis is the development of a library of extended collective functions, which complements and improves the existing UPC standard library with programmable constructs based on efficient algorithms. A UPC MapReduce framework (UPC-MR) has also been implemented to support this highly scalable computing model for UPC applications. Finally, the analysis and development of relevant kernels and applications (e.g., a large parallel particle simulation based on Brownian dynamics) confirm the usability of these libraries, concluding that UPC can provide high performance and scalability, especially for environments with a large number of threads at a competitive development cost