Service elastic expansion and contraction method based on emergency detection

Abstract

The invention discloses a service elastic expansion and contraction method based on emergency detection. The method comprises the following steps: taking an obtained service state as a work cycle; obtaining work load data before the service, and inputting the work load data into the long-term prediction model to carry out multi-step probability prediction to obtain prediction results of a plurality of time points in the future; calculating a multi-step deviation vector; updating a confidence interval upper bound value of the multi-step probability prediction result; if at least two values in the multi-step deviation vector are greater than a preset threshold value, determining that the current state is an emergency; otherwise, the emergency situation is not present; generating a trend prediction result; calculating the number of instances required by the service at the moment according to the preset workload upper limit of a single instance; obtaining the workload data of the previous service from the time sequence database, and carrying out single-step prediction to obtain a single-step prediction result; obtaining a future instance number of the service based on a base point offset method; and scheduling is carried out, so that a work cycle is completed.

Description

technical field [0001] The invention relates to the technical field of cloud computing, and more particularly, to a service elastic scaling method based on emergency detection. Background technique [0002] The existing elastic scaling methods of services can be divided into reactive scaling methods and predictive scaling methods: [0003] Among them, responsive scaling methods include threshold method and reinforcement learning-based method. This method will only scale the instance after getting the changes of the current service. Because the workload is changing all the time, the previous scaling may not be accurate enough after completion. In addition, the number of instances that can be scaled at a time is limited by the preset maximum number of instances. When the workload changes more than expected, the number of scaled instances will not keep up. [0004] The predictive method uses a method based on time series prediction to model and analyze the workload, so as to...

AI Technical Summary

Problems solved by technology

Therefore, this method will not achieve optimal scaling results

[0007] Second, the prior art method does not analyze and process the prediction error of the workload prediction model

However, in practice, if the differential integrated moving average autoregressive model (ARIMA) is used, there must be prediction errors, such as the general predictive scaling method, which will deal with this part of the error

However, this method cannot perform corresponding processing, and the elastic scaling strategy determined on this basis must be inaccurate

[0008] Third, the existing technology does not perform well for workloads with large fluctuations

[0010] However, the scaling of the responsive method is limited by the preset maximum number of scaling instances, which may not be able to catch up with the changing speed of the workload, and the response speed is slow; at the same time, the prediction results of the predictive method based on the time series prediction method may be There will be a large error, and the scaling strategy based on this is unstable

Images