A Time Series Forecasting Approach to Minimize Cold Start Time in Cloud-Serverless Platform

TL;DR

This paper proposes a time series forecasting method using SARIMA to predict request arrivals in serverless platforms, enabling proactive autoscaling to reduce cold start latency and resource wastage.

Contribution

It introduces a novel prediction-based autoscaling approach that improves cold start times and resource efficiency in serverless computing environments.

Findings

PBA outperforms default HPA in reducing cold start latency.

The approach decreases resource wastage during scaling.

Forecasting request arrivals improves autoscaling responsiveness.

Abstract

Serverless computing is a buzzword that is being used commonly in the world of technology and among developers and businesses. Using the Function-as-a-Service (FaaS) model of serverless, one can easily deploy their applications to the cloud and go live in a matter of days, it facilitates the developers to focus on their core business logic and the backend process such as managing the infrastructure, scaling of the application, updation of software and other dependencies is handled by the Cloud Service Provider. One of the features of serverless computing is ability to scale the containers to zero, which results in a problem called cold start. The challenging part is to reduce the cold start latency without the consumption of extra resources. In this paper, we use SARIMA (Seasonal Auto Regressive Integrated Moving Average), one of the classical time series forecasting models to predict the time at which the incoming request comes, and accordingly increase or decrease the amount of required containers to minimize the resource wastage, thus reducing the function launching time. Finally, we implement PBA (Prediction Based Autoscaler) and compare it with the default HPA (Horizontal Pod Autoscaler), which comes inbuilt with kubernetes. The results showed that PBA performs fairly better than the default HPA, while reducing the wastage of resources.