A traceability method for gpu computing performance problems

Abstract

The invention provides a tracing method for a GPU computing performance problem. The method comprises the following steps: (1) describing behavior characteristics of a CUDA program kernel function level by using a CUDA task parallel model CTPM, thereby identifying a performance low-efficiency behavior of the CUDA program kernel function level, and expanding kernel function level performance optimization work; and (2) for the internal data parallel part of the CUDA program kernel function, using a hierarchical iteration attribute reduction algorithm LIAR to identify the key metric, performing deep performance mining on the CUDA program, and optimizing the internal performance problem of the CUDA kernel function. According to the method, by analyzing the CUDA development operation environment and the programming model based on the GPU hardware characteristics, an effective performance problem tracing method is found, the accuracy of performance bottleneck positioning is improved, and thepurpose of shortening the optimization period is achieved.

Description

technical field [0001] The invention relates to the field of performance monitoring and analysis in the field of information technology, in particular to a source tracing method for GPU computing performance problems. Background technique [0002] In recent years, the development of high-performance computers is changing with each passing day. The traditional single computing structure is gradually replaced by a hybrid computing structure, in order to give full play to the strengths of various components and play an effective role in improving computing capabilities. In the hybrid structure, GPU plays an increasingly important role as the acceleration part of the current hybrid structure. At the same time, the mixed computing mode increases the complexity of programming. [0003] Currently, NVIDIA's GPU and its development and operating environment and programming model CUDA are the most mature and widely used. Usually, the CUDAkernel function-level program analysis method...

AI Technical Summary

Problems solved by technology

With this execution mode, it is more difficult to use the behavior description method to analyze performance and trace the source of performance problems

[0004] In addition, calculating metrics with statistical meaning is currently the main method for internal analysis of kernel functions. However, the structure of the GPU is complex, and there are usually many metrics for computing performance, including memory operations, computing operations, branches, and synchronization operations. These are all related to performance. More or less relationship, if each item is optimized separately, the workload will be very large, these metrics are usually interrelated, and it is often impossible to achieve full optimization

Images