A cascade fault-tolerant processing method adapted to cpu and gpu heterogeneous clusters

Abstract

The invention relates to a cascading fault-tolerant processing method adaptive to a CPU and GPU heterogeneous cluster. The method comprises the steps that a data transmission consistency detection model is constructed and used for detecting consistency of data transmission; a data access consistency model is constructed and used for realizing consistency of data access between a CPU and a GPU; a data operation result correctness detection model is constructed and used for detecting the correctness of a thread data operation result; a service backup model is constructed on an application layer and used for backing up historical records of service running; and a service job information backup model is constructed on a system layer and used for backing up job information of service running. In this way, when a non-physical damage fault occurs to the CPU and GPU heterogeneous cluster, a service fault is quickly positioned for business personnel, a pre-fault state is extracted, a service is reset quickly, and losses are reduced.

Description

technical field [0001] The present disclosure relates to the field of computer technology, and in particular to a cascade fault-tolerant processing method adapted to CPU and GPU heterogeneous clusters. Background technique [0002] Since the birth of the first computer, computer technology has developed rapidly, ranging from personal computers to supercomputers in various forms. Human’s demand for computing is endless. In order to meet human’s demand for computing power, the speed of computers produced is getting faster and faster. Occupying the first and second places in the world's supercomputer rankings, the operating capacity reached 100Pbit (petabit, gigabit) for the first time, reflecting the comprehensive capabilities of our country. But at the same time, supercomputing capability is accompanied by a super complex and huge computer architecture. For example, the main computing capability of "Tianhe-2" is provided by GPU (Graphics Processing Unit, Graphics Processing ...

AI Technical Summary

Problems solved by technology

Large-scale data analysis and processing applications have high computational complexity, use large-scale computing resources, run for a long time, and require high reliability of computing resources. Some application services have a large amount of data access and require large-scale read-write storage devices; some Communication is intensive and requires high network transmission of the system; some occupy a large amount of GPU space, frequent read and write operations, and require high GPU stability; some processes are complex, with many intermediate links, and single-step errors affect the entire service process. High fault tolerance requirements

Images