Performance evaluation method and system for GPU applications in CPU-GPU heterogeneous environment

Abstract

The invention discloses a performance evaluation method and system for GPU applications in a CPU-GPU heterogeneous environment and belongs to the field of GPU performance evaluation. Specifically, themethod comprises the steps of learning performance condition of various applications operated on GPU architecture based on a decision-making tree algorithm in machine learning and establishing a decision-making tree model; acquiring monitoring characteristics of which influences on application performance time are the highest in sequence within a decision-making tree matching process, namely ordering the important degree of the characteristics; carrying out correspondence on screened characteristic sets and four common problems of the applications in sequence, wherein the four common problemsmainly indicate computing correlation, memory correlation, occupation rate correlation and synchronization correlation, thereby preliminarily obtaining problem directions in which performance bottlenecks of to-be-analyzed applications are located. According to the method and the system, through utilization of the method of combining the decision-making tree model with analysis modeling, the universal, relatively accurate, rapid, simple and easy-to-use method for carrying out the performance evaluation on resources and applications on a GPU is provided.

Description

technical field [0001] The invention belongs to the field of GPU performance evaluation, and more specifically relates to a method and system for performance evaluation of resources and applications on the GPU in combination with machine learning and analytical modeling in a CPU+GPU hybrid heterogeneous environment. Background technique [0002] With the continuous development of science and technology, all aspects put forward higher requirements for high-performance computing. GPU has powerful computing power, high memory bandwidth, low power consumption, and good programmability, but it does not handle logic well, making CPU-GPU heterogeneity an inevitable trend. However, although CPU-GPU heterogeneous computing nodes can achieve high performance, the actual performance is often not ideal, and computing resources and storage bandwidth cannot be effectively utilized. The reason for performance degradation lies in many aspects, including load imbalance caused by uneven task...

AI Technical Summary

Problems solved by technology

[0004] However, there are certain defects in the above method

Although the traditional performance analysis model has high accuracy, it requires a detailed understanding of hardware architecture knowledge. The methods of data acquisition and modeling are very complicated, often time-consuming, and the model is difficult to use; or for a specific Built for architecture or application, not universal

Although the method based on machine learning is simple and easy to use, its accuracy strongly depends on the training data set and the selection of feature values ​​by the model itself; and there are few studies on performance evaluation using this method, which is mainly used for performance prediction. Cannot reflect performance bottlenecks and guide application optimization

Images