GPU (Graphics Processing Unit) task scheduling method and system

Abstract

The invention provides a GPU (Graphics Processing Unit) task scheduling method and system. The system comprises an application analysis module and a dynamic task scheduling module, wherein the application analysis module is used for obtaining the instruction number of each kernel of an application program; the dynamic task scheduling module is used for judging whether the amount of kernels which operate at present achieves a preset upper limit value or not; if the amount of kernels which operate at present does not achieve the preset upper limit value, the kernel which forms a highest kernel combination priority with the kernel which operates at present from seized kernel and new arrival kernel; if the amount of kernels which operate at present achieves the preset upper limit value, whether a kernel which can be combined with the kernel which operates at present to obtain a kernel combination with a higher priority is in the presence in the seized kernel and the new arrival kernel or not is continuously judged; if the kernel is in the presence, whether GPU performance improvement after a program is seized is greater than GPU expenditures occupied in a seizing process or not is judged; and if the GPU performance improvement after the program is seized is greater than the GPU expenditures occupied in the seizing process, seizing is carried out. By use of the system, delay is effectively lowered, a resource use ratio is improved, and the performance of the GPU is improved.

Description

technical field [0001] The present invention relates to the technical field of processors, in particular to the technical field of GPUs, and specifically relates to a GPU task scheduling method and system. Background technique [0002] With the rapid growth of Internet users and the rapid expansion of data volume, the demand for computing in data centers is also rising rapidly. The emergence of various new applications, such as deep learning online prediction, video transcoding, and image compression and decompression, has required computing requirements that have far exceeded the capabilities of traditional CPU processors. The emergence of new processors such as GPUs that are different from CPUs has brought about huge architectural changes in data centers. [0003] Traditional high-performance computing clusters use MPI and other methods as the main communication method, and use the queue system to allow each task to exclusively enjoy the resources in the cluster. Compare...

AI Technical Summary

Problems solved by technology

[0016] In view of the shortcomings of the prior art described above, the purpose of the present invention is to provide a GPU task scheduling method and system for solving the problems of GPU task scheduling that affect memory performance, high delay, and low resource utilization in the prior art

Images