Detecting and Handling Flash-Crowd Events on Cloud Environments

TL;DR

This paper introduces a novel entropy-based detection mechanism for flash-crowd events in cloud environments, formulates the handling as an integer programming problem, and proposes an algorithm to optimize resource allocation during such events.

Contribution

It presents a new detection method using entropy and total correlation, formulates the handling as an integer programming problem, and introduces the FCHP-ILS algorithm for resource management.

Findings

The entropy-based detection method effectively identifies flash-crowd events.

FCHP-ILS improves resource allocation during flash crowds compared to standard auto-scaling.

Case study shows the proposed approach outperforms Amazon's Auto Scaling in simulated scenarios.

Abstract

Cloud computing is a highly scalable computing paradigm where resources are delivered to users on demand via Internet. There are several areas that can benefit from cloud computing and one in special is gaining much attention: the flash-crowd handling. Flash-crowd events happen when servers are unable to handle the volume of requests for a specific content (or a set of contents) that actually reach it, thus causing some requests to be denied. For the handling of flash-crowd events in Web applications, clouds can offer elastic computing and storage capacity during these events in order to process all requests. However, it is important that flash-crowd events are quickly detected and the amount of resources to be instantiated during flash crowds is correctly estimated. In this paper, a new mechanism for detection of flash crowds based on concepts of entropy and total correlation is proposed. Moreover, the Flash-Crowd Handling Problem (FCHP) is precisely defined and formulated as an integer programming problem. A new algorithm for solving it, named FCHP-ILS, is also proposed. With FCHP-ILS the Web provider is able to replicate contents in the available resources and define the types and amount of resources to instantiate in the cloud during a flash-crowd event. Finally we present a case study, based on a synthetic dataset representing flash-crowd events in small scenarios aiming at comparing the proposed approach with de facto standard Amazon's Auto Scaling mechanism.