Data Centric Programming Abstractions Suitable from a Multiphysics Application Perspective
Programming Models & Languages
Scientific Software Development
TimeWednesday, June 27th1:45pm - 2:07pm
DescriptionOrchestrating locality in data management has always been an important concern in high performance computing (HPC) systems. In distributed memory architectures the cost variation in data access had a few layers; cache hierarchy and DRAM access versus off node. In forthcoming architectures there are many more cost variants that can almost be viewed as groups close in locality, separated by jumps in locality loss and corresponding costs. Data locality has not traditionally been an expressive feature of languages used in HPC applications for various reasons, however, with the increasing complexity of the memory hierarchy and higher parallelism in emerging HPC systems, developers can no longer limit themselves to low-level solutions and ignore locality abstractions. Data locality abstractions are emerging in the forms of libraries, data structures, languages and runtime systems; a common theme is increasing productivity without sacrificing performance. In this presentation I will summarize trends and commonalities that emerged from a series of workshops, PADAL, for exploration of data locality abstractions. We found that various locality concepts can be combined to develop a comprehensive approach to expressing and managing data locality on future large-scale high-performance computing systems.