Browsing by Author "Schuchart, Joseph Konstantin"

Filter results by typing the first few letters
Now showing 1 - 1 of 1
  • Thumbnail Image
    ItemOpen Access
    Global task data dependencies in the partitioned global address space
    (Stuttgart : Höchstleistungsrechenzentrum, Universität Stuttgart, 2021) Schuchart, Joseph Konstantin; Resch, Michael M. (Prof. Dr.-Ing. Dr. h.c. Dr. h.c. Prof. E.h.)
    High-Performance Computing (HPC) has become an important part of scientific discovery in many fields and takes an important role in many engineering processes, harnessing the power of large amounts of computational resources to gain insights into otherwise hidden technological and natural phenomena. The dominating programming model driving today’s parallel applications is a two-level approach consisting of message-based communication between processes using MPI and static loop-level thread-parallel execution using OpenMP constructs. However, two programming models have tried to challenge this status quo. First, the Partitioned Global Address Space (PGAS) model is an attempt to elevate shared memory programming to the level of distributed systems and to directly expose modern network hardware features to the application developer. Second, task-based programming aims at providing abstractions that help discover a greater amount of concurrency in parallel applications, which in turn can be used to better exploit the computational resources at hand. Both models are an attempt to break up the strict synchronization imposed by the traditional models: the PGAS model decouples synchronization and communication while task-based programming models minimize the required synchronization to a set of constraints on the order of the execution of tasks. This work proposes a novel way of orchestrating the execution of tasks at a global scale by using distributed task graph discovery and data dependencies in the global memory space. The results demonstrate that applications exhibiting concurrency beyond single loop parallelism may use this new model to significantly improve performance and scalability by combining the benefits of task-based programming and one-sided communication in the PGAS model.