Trade-off Analysis in Learning-augmented Algorithms with Societal Design Criteria

Abstract

Traditionally, computer systems are designed to optimize classic notions of performance such as flow completion time, cost, etc. The system performance is then typically evaluated by characterizing theoretical bounds in worst-case settings over a single performance metric. In the next generation of computer systems, societal design criteria, such as carbon awareness and fairness, becomes a first-class design goal. However, the classic performance metrics may conflict with societal criteria. Foundational understanding and performance evaluations of systems with these inherent trade-offs lead to novel research questions that could be considered new educational components for performance analysis courses. The classic techniques, e.g., worst-case analysis, for systems with conflicting objectives may lead to the impossibility of results. However, a foundational understanding of the impossibility of results calls for new techniques and tools. In traditional performance evaluation, to understand the foundational limits, typically, it is sufficient to derive lower-bound arguments in worst-case settings. In the new era of system design, lower bounds are inherently about trade-offs between different objectives. Characterizing these trade-offs in settings with multiple design criteria is closer to the notion of Pareto-optimality, which is drastically different from classic lower bounds. With the impossibility of results using classic paradigms, one possible direction is to (re)design systems following the emerging direction of learning-augmented algorithms. With this approach, it might be possible to remove/mitigate the foundational conflict between classic vs. societal metrics using the right predictions. However, the performance evaluation of learning-augmented algorithms calls for a new set of technical questions, which we highlight in this paper.