Multistage nested data drive calculation method matched with high-performance computer structure

Abstract

The invention provides a multistage nested data drive calculation method matched with a high-performance computer structure. The method is mainly used for adapting to the high-performance computer system structure. The method comprises the first step of three-stage nested parallel calculation organization of courses, thread groups and threads matched with calculation resources namely calculation nodes, processors and processor cores, wherein the static data distribution method is adopted in the two stages of the courses and the thread groups and the thread groups support the task dynamic dispatching; the second step of three-stage nested data communication and data partitioning matched with the network, an internal storage and a cache. The method solves the problem that when data drive parallel calculation is implemented in the prior art, calculation parallelism and data locality are low in efficiency, and achieves the technical effects of effectively balancing loads and improving communication, internal storage access and calculation efficiency.

Description

technical field [0001] The invention relates to the field of computer technology, in particular to a multi-level nested data-driven calculation method matching the structure of a high-performance computer. Background technique [0002] The mainstream high-performance computer architecture is a distributed multi-core multi-processor cluster, and its development trend is the increase of computing resources and the deepening of storage levels. Specifically, the organization level of computing resources is: cluster--computing nodes--processor--processor core, and the organization level of corresponding storage resources is: distributed memory--local shared memory--the outermost shared memory of the processor Cache--Processor core private Cache, the actual performance of the current high-performance computer system depends on the degree of parallelism of the program and the locality of data access. [0003] In high-performance computing, data-driven (Data-Driven) parallel comput...

AI Technical Summary

Problems solved by technology

[0009] Embodiments of the present invention provide a multi-level nested data-driven computing method that matches the structure of a high-performance computer to solve the inefficiencies in computing parallelism and data locality faced in the prior art when implementing data-driven parallel computing technical issues, the method includes:

Images