Modeling and predicting measured response time of cloud-based web services using long-memory time series

TL;DR

This study analyzes the response time of cloud services, demonstrating that long-memory models like ARFIMA significantly improve forecast accuracy over traditional short-memory models, aiding better service selection.

Contribution

The paper introduces the application of long-memory ARFIMA models to predict cloud service response times, showing substantial error reduction compared to ARIMA models.

Findings

ARFIMA models reduce forecast error by up to 57.8% compared to ARIMA.

Long memory is statistically verified in CPU-intensive cloud services.

Considering long-range dependence improves QoS prediction accuracy.

Abstract

Predicting cloud performance from user's perspective is a complex task, because of several factors involved in providing the service to the consumer. In this work, the response time of 10 real-world services is analyzed. We have observed long memory in terms of the measured response time of the CPU-intensive services and statistically verified this observation using estimators of the Hurst exponent. Then, na\"ive, mean, autoregressive integrated moving average (ARIMA) and autoregressive fractionally integrated moving average (ARFIMA) methods are used to forecast the future values of quality of service (QoS) at runtime. Results of the cross-validation over the 10 datasets show that the long-memory ARFIMA model provides the mean of 37.5 % and the maximum of 57.8 % reduction in the forecast error when compared to the short-memory ARIMA model according to the standard error measure of mean absolute percentage error. Our work implies that consideration of the long-range dependence in QoS data can help to improve the selection of services according to their possible future QoS values.