EURO 2024 Copenhagen
Abstract Submission

EURO-Online login

322. Optimal is the enemy of good: Solving very large scale virtual machine placement problems at Amazon Elastic Compute Cloud

Invited abstract in session MC-1: Ruben Ruiz, stream Keynotes.

Monday, 12:30-14:00
Room: Sportshallen (building: 101)

Authors (first author is the speaker)

1. Rubén Ruiz
Departamento de Estadistica e Investigación Operativa Aplicadas y Calidad, Universitat Politècnica de València

Abstract

A significant portion of the effort within the field of Operations Research is dedicated to the development of intricate ad-hoc algorithms and models tailored for addressing diverse optimization problems. Frequently, the research community exhibits a preference for complexity in the proposed methodologies. Notably, an esteemed characteristic sought in these methods is the incorporation of problem-specific knowledge, enabling the attainment of results that get as close as possible to optimality.

In academic publishing, the pursuit of marginal gains in optimality gaps is a common goal, even at the expense of introducing additional complexity to models or algorithms. While there is consensus that such endeavors contribute to advancements in algorithms and propel the field of Operations Research forward, a frequently underestimated dimension is the practical applicability of these advancements in industrial settings.

Real industrial problems are often messy with imprecise (or numerous) objectives, soft constraints, preferences and a myriad of situations that demand pragmatic approaches. Additionally, these problems undergo rapid evolution with frequent model refinements, occurring often fortnightly. Here, the ongoing painful adaptation of intricate code bases stemming from ad-hoc methods is not the favored choice. In contrast, solvers offer greater flexibility, allowing faster implementation of new constraints through their APIs than altering tailored algorithms. In industrial contexts, the preference lies in flexibility, maintainability, robustness, and reduced operational load, where a modest optimality gap is deemed a minor trade-off.

Moreover, in the realm of large-scale real-world problems, mathematical solvers often prove impractical. It is commonplace to resort to simplifying the problem for solvability. This practice raises a critical question: Is it logical to insist on chasing optimality in solving a simplified problem with a reduced real-world fidelity? Furthermore, real problems entail vast datasets, often including forecasted or approximated input data. Does it make sense to go great lengths to optimally solve a problem when a portion of the input data involves approximations?

In this presentation, I advocate for relinquishing the pursuit of optimality and, instead, embracing heuristic solvers and simplified modeling approaches. Such strategies yield expedient, adaptable, maintainable, and easily implementable models. The discourse will draw upon various examples, spanning classical routing and scheduling problems, culminating in intricate real-world virtual machine placement bin packing problems encountered at Amazon Elastic Compute Cloud (EC2).

Keywords

Status: accepted

Back to the list of papers