Optimization Heuristics for Cost-Efficient Long-Term Cloud Portfolio Allocations Under Uncertainty

TL;DR

This paper addresses the complex problem of cost-efficient cloud resource allocation by proposing formal models and two heuristics, demonstrating that a simple first-fit approach outperforms genetic algorithms in speed and quality.

Contribution

It introduces a formal specification for cloud portfolio management considering market types and compares two heuristics for stochastic resource allocation.

Findings

First-fit heuristic outperforms genetic algorithm in speed and solution quality.

The formal model captures the nuances of cloud market types.

Heuristics effectively address the stochastic temporal bin packing problem.

Abstract

Today's cloud infrastructure landscape offers a broad range of services to build and operate software applications. The myriad of options, however, has also brought along a new layer of complexity. When it comes to procuring cloud computing resources, consumers can purchase their virtual machines from different providers on different marketspaces to form so called cloud portfolios: a bundle of virtual machines whereby the virtual machines have different technical characteristics and pricing mechanisms. Thus, selecting the right server instances for a given set of applications such that the allocations are cost efficient is a non-trivial task. In this paper we propose a formal specification of the cloud portfolio management problem that takes an application-driven approach and incorporates the nuances of the commonly encountered reserved, on-demand and spot market types. We present two distinct cost optimization heuristics for this stochastic temporal bin packing problem, one taking a naive first fit strategy, while the other is built on the concepts of genetic algorithms. The results of the evaluation show that the former optimization approach significantly outperforms the latter, both in terms of execution speeds and solution quality.