Task scheduling system based on smart processors

Abstract

The invention discloses a task scheduling system based on smart processors. The task scheduling system comprises a user layer, a system layer and a hardware layer. The user layer comprises a user interface library. The system layer comprises a communication module, a task preprocessing module, a task queue, a load balancing module and a task scheduling module. The hardware layer comprises task cache queues, application dictionaries, communication modules and a plurality of dockers of smart processor. According to the task scheduling system based on the smart processors provided by the invention, under a standalone multi-smart processor condition, the programming difficulty is simplified, the application deployment is simplified, the application isolation is realized, the automatic scheduling is realized, and a utilization rate of the smart processors is maximized. The consequence that the throughput of the system is relatively low, the corresponding time of tasks is relatively long and the user experience is relatively low is avoided.

Description

technical field [0001] The invention relates to computer information processing, in particular to a task scheduling system based on an intelligent processor. Background technique [0002] As deep learning technology has made breakthroughs in the fields of image classification, language recognition, object classification, and machine translation, such as: in the 2012 ImageNet competition, Alex Krizhevshy and others used convolutional neural networks to achieve amazing results; deep learning systems AlphaGo defeated Korean chess player Lee Sedol with a total score of 4:1; the neural network-based machine translation system developed by Google can greatly improve the accuracy of translation. Deep learning technology is increasingly valued by academia and industry. [0003] However, the scale of the network increases with the complexity of the task, and the amount of calculation required also increases sharply. For example, the VGG16 network requires 31G floating-point calculat...

AI Technical Summary

Problems solved by technology

In the case of a single machine with multiple cards, tasks have more scheduling options. At the same time, problems such as load balancing are also introduced. Task scheduling and management become more complicated, and it becomes more difficult to maximize the utilization of intelligent processors.

It will eventually lead to serious consequences such as low system throughput, long task response time, and poor user experience.

Images