Coflow scheduling method based on deep reinforcement learning

Abstract

The invention discloses a Coflow scheduling method based on deep reinforcement learning. Based on deep reinforcement learning and in combination with a multi-level priority queue and a basic flow scheduling algorithm, a Markov decision model is established, a Coflow scheduling strategy is expressed in a formalized mode, and then adjustment of the Coflow scheduling strategy is completed. Accordingto the method, a state matrix is formed by using the data size, the total flow number, the active flow number and the duration of Coflow in the current network environment to represent the network state, define and adjust the action of a flow transmission multistage priority queue, an evaluation equation is established to realize the minimum optimization of the Coflow average completion time. According to the method, deep reinforcement learning and multi-level priority queues are combined, the Coflow scheduling strategy is adjusted through a large number of trial and error and learning in theCoflow sending scheduling environment, more environmental parameters are brought into account compared with a traditional heuristic algorithm, and the transmission efficiency of Coflow is effectivelyimproved.

Description

technical field [0001] The present invention relates to Coflow scheduling, and more specifically, to a Coflow scheduling method based on deep reinforcement learning. Background technique [0002] Big data computing frameworks headed by Apache Spark are often used in computing tasks such as data mining and machine learning. They use a distributed architecture to divide the input data and assign the divided small data blocks to different computing nodes for calculation. , after the data processing is completed, it will be redistributed and calculated. After one or more iterations, the desired result will be finally obtained. With the execution of the above computing applications, a large number of communication flows for transmitting data will be generated between different nodes (Nodes) in the cluster. These communication flows are highly parallel and occupy most of the application running process, profoundly affecting the calculation process of big data applications. There...

AI Technical Summary

Problems solved by technology

The multi-level feedback queue has the advantages of small scheduling delay and can effectively distinguish between long flow and short flow, but there are still defects, that is, the threshold value often needs to be manually set in advance, which is empirical, and the network environment is complex and changes frequently. Thresholds are often not optimal, which limits the final effect of scheduling

Images