Scaling HabaneroUPC++ on Heterogeneous Supercomputers
Vivek Kumar, Max Grossman, Hongzhang Shan, Vivek Sarkar · 2015
Accelerators/co-processors have made their way into supercomputing systems. These modern heterogeneous systems feature multiple layers of memory hierarchies, and produce a high degree of thread-level parallelism. To ensure that current and future applications perform well on these systems, it is important that users be able to cleanly express the various types of parallelism found in their applications while trusting that expertly-implemented runtime libraries will schedule this parallelism in a way that efficiently utilizes the memory hierarchies and computational resources of their system. In this poster we present our work-in-progress, a distributed, heterogeneous programming model in HabaneroUPC++, which aims to target distributed, heterogeneous systems with multi-layered memory hierarchies through a distributed data-driven programming model that integrates all memory layers together through data-flow programming.