Synchronous programming models make such decisions exceedingly difficult, if not impossible.Ĭonsequently, a programming paradigm built around a far less synchronous approach is needed. Applications ready for exascale feature such potential for parallelism, but once exposed, decisions must be made at runtime for the efficient placement of work. Extracting performance from such hardware, where the level of parallelism is increased 1,000 fold, requires that one exploits an even larger degree of parallelism from the application.
To aggregate a collection of computing elements that is growing exponentially, new system architectures rely on hierarchical composition: the memory hierarchy deepens, thereby increasing non-uniform memory access (NUMA) effects network interconnects use high-dimension tori, or other scalable structures and manycore CPUs are embedded in the computing nodes, either through the use of separate or self-hosted accelerators. These limitations require application developers to make a significant development shift to transition their applications toward modern programming paradigms. The limitations of current programming paradigms-relative to the new concurrency landscape-are well known, and the search for alternative approaches has demonstrated proof-of-concept impacts in application scalability and efficiency. Our solution, implemented in production form in the PaRSEC environment, focuses on developing a programming paradigm that can both expose and dynamically manage massive levels of parallelism while delivering performance portability across heterogenous systems. Moreover, the fact that the amount of available parallelism will vary dynamically means that it is problematic to rely on known techniques, such as MPI+X, to consistently exploit all of the potential concurrency. Programming models that cannot expose and exploit such vast amounts of parallelism, even as it varies dynamically in response to ongoing system conditions, will struggle to reach exascale or even near exascale. It is widely agreed that this unprecedented increase in concurrency-at least three orders of magnitude-is among the most formidable challenges of extreme-scale computing. WHY?Įxascale computing will require many system architecture changes, including a drastic increase in hardware parallelism. The Distributed Tasking for Exascale (DTE) project plans to further develop these PaRSEC features for the Exascale Computing Project (ECP)-in terms of scalability, interoperability, and productivity-to line up with the critical needs of ECP application communities. These features comprise a development framework that supports multiple domain-specific languages and extensions and includes tools for debugging, trace collection, and analysis. The Parallel Runtime Scheduling and Execution Controller (PaRSEC) environment provides a runtime component capable of dynamically executing on heterogeneous distributed systems along with a productivity toolbox.
0 kommentar(er)
