Data center network energy consumption and service quality optimization method based on reinforcement learning

Abstract

The invention discloses a data center network energy consumption and service quality optimization method based on reinforcement learning. The method comprises the steps of constructing data center network energy consumption and service quality optimization model based on a deep reinforcement learning framework; respectively taking the link utilization rate, the completion time related performancecalculated by the network performance and the link margin ratio as the state, reward and action of the optimization model; and then adjusting the flow of the data center network according to the linkmargin ratio output by the optimization model, so that the time volatility of the data flow and the spatial distribution characteristic of the data flow are both considered in the adjustment process,and the energy efficiency of the data center network is improved while the FCT is ensured.

Description

technical field [0001] The invention belongs to the technical field of computer networks, and in particular relates to a data center network energy consumption and service quality optimization method based on reinforcement learning. Background technique [0002] High power consumption in data centers has become a major problem for data center operators. Recent studies show that electricity consumption by U.S. data centers is expected to reach 139 billion kWh in 2020. In a data center, a data center network (DCN) consisting of switches and links consumes about 10% to 20% of the total energy consumption. Traffic in DCNs can generally be divided into two categories: delay-sensitive traffic and delay-tolerant traffic, where delay-sensitive traffic is mainly traffic for delay-sensitive services (such as web search), which is generally small (from a few KB to MB) , and has a clear data flow completion time (FCT) limit; and delay tolerant traffic is usually large (hundreds of MB ...

AI Technical Summary

Problems solved by technology

[0004] However, the existing traffic scheduling schemes have the following two deficiencies: First, rough integration schemes may lead to service QoS degradation

Second, the existing solutions do not fully consider the diversity of DCN traffic types, and the adaptability is poor

However, using these two features to obtain an optimized merging scheme results in high computational complexity

Most of the existing schemes only give some fixed models, which cannot be well extended to most situations