Novel Methods to Resolve the Data Analysis Bottleneck in GPU-accelerated Stencil Computations

The NEOSC project studies two complementary methods to resolve the data analysis bottleneck in GPU-accelerated stencil computations: a data-driven reduced initial-condition (RIC) algorithm and an opportunistic data operations platform (ODOP) that need to work in symbiosis to enable generic GPU-accelerated iterative stencil loop (ISLs) in exascale. 

ISLs form a class of algorithms, in which the elements of a computational grid are updated by sampling their neighborhood in a fixed pattern that resembles a stencil.

ISLs are usually implemented as computations on structured grids, which have been identified as one of the major recurring computational patterns in high-performance computing (HPC). However, attempts to include data analysis tasks to the GPU-accelerated algorithm often affects the performance, as these operations are global in nature, requiring too many memory transactions. For this reason, data analysis tasks are commonly performed in post-processing. Changing algorithms to allow the execution of data analysis tasks opportunistically along the stencil operations, in other words opportunistic data operations, can substantially improve the performance, data quality and cost-efficiency. 

The RIC algorithm and ODOP address the weaknesses in the current use of large-scale computers and therefore show potential for speeding up simulations based on iterative stencil loops on current as well as future exascale systems.

CSC HPC environments are to be used throughout the NEOSC project. CSC will be involved in identifying and describing operational patterns occurring during the execution of the HPCGP, integrating the defined data operations and their descriptors into ODOP tools as well as implementing a proof-of-concept HPCGP-ODOP-RIC pipeline. In addition, CSC will contribute to the development ODOPRuntime and the integration of ODOPRuntime into supercomputing systems.

Funded by the European Union logo.