Data Access Method for Distributed Heterogeneous Computing

Abstract

The disclosure provides a data access method for distributed heterogeneous computing, including a method for organizing, managing and acquiring global data and a method for local data. Among them, the globally consistent data organization method and the management method of unified network shared data and local shared data enable a certain data consumption computing node in a distributed heterogeneous system to transparently obtain network information by name from a data production node or a networked data management system. Shared data; it also enables each computing component inside the computing node to obtain local shared data by name. Since different heterogeneous computing devices and different computing nodes have different expressions for the same data structure, the data management method and system provide transparent data conversion operations at the same time to adapt to each computing device in a distributed heterogeneous computing system. Different requirements for data format and representation.

Description

technical field [0001] The present disclosure relates to the technical field of distributed heterogeneous computing, and in particular to a data access method for distributed heterogeneous computing, specifically including a method for managing global data sets in a distributed heterogeneous computing system and a method for managing local datasets in a distributed heterogeneous computing system. method of the dataset. Background technique [0002] In recent years, heterogeneous fusion parallel computing systems have gradually become the first choice for high-performance and low-power computing. Inspired by the success of GPU general-purpose computing and the failure of CPU Moore's Law, modern computing often integrates a variety of computing devices, such as for Artificial intelligence tensor computing unit (TPU), dedicated digital signal processor (DSP) and field programmable gate array (FPGA). These different computing hardware form heterogeneous computing units through ...

AI Technical Summary

Problems solved by technology

Since each type of computing hardware supports different data structures, and at the same time, different data objects of computing tasks are distributed in different computing units of the computing system, thus causing difficulties in data management of this type of computing system

Traditional parallel computing systems use protocols such as Message Passing Interface (MPI) for communication. This type of protocol was generally defined in the 1990s and works in CPU homogeneous parallel computing systems, regardless of the difference between computing units and computing hardware. In addition, MPI Such protocols are mainly aimed at large-scale numerical calculations, which support fewer data types, and are gradually unsuitable for modern heterogeneous fusion parallel computing systems.

Images