The authors study an online data replication problem in a distributed system, where the goal is to dynamically create and delete data copies to minimize the total storage and network cost of serving a sequence of data access requests. They consider the learning-augmented setting, assuming simple binary predictions about inter-request times at individual servers.
The key highlights of the paper are:
The authors propose an online algorithm that integrates and balances between following predictions and not using predictions. They introduce a hyper-parameter α to represent the level of distrust in the predictions.
Theoretical analysis shows that the proposed algorithm is 5+α/3-consistent (competitive ratio under perfect predictions) and 1+1/α-robust (competitive ratio under terrible predictions).
The authors establish a lower bound of 3/2 on the consistency of any deterministic learning-augmented algorithm for this problem, implying that it is not possible to achieve a consistency approaching 1.
Experimental evaluations using real data access traces demonstrate that the algorithm can make effective use of predictions to improve performance with increasing prediction accuracy.
A otro idioma
del contenido fuente
arxiv.org
Ideas clave extraídas de
by Tianyu Zuo,X... a las arxiv.org 04-26-2024
https://arxiv.org/pdf/2404.16489.pdfConsultas más profundas