Price of asynchrony in mobile agents computing

  • Authors:
  • Yoann Dieudonné;Andrzej Pelc

  • Affiliations:
  • MIS, Université de Picardie Jules Verne Amiens, France;Département dinformatique, Université du Québec en Outaouais, Gatineau, Québec J8X 3X7, Canada

  • Venue:
  • Theoretical Computer Science
  • Year:
  • 2014

Quantified Score

Hi-index 5.23

Visualization

Abstract

Asynchrony is one of the main challenges in distributed computing. Some tasks, such as distributed Byzantine consensus, are impossible in the asynchronous setting, while they can be carried out synchronously. For other tasks, such as rendezvous in arbitrary graphs, the best known synchronous algorithm has cost much lower than the best asynchronous one. Various degrees of asynchrony and synchrony and comparisons between them in terms of feasibility of distributed tasks have been particularly intensely studied in the context of mobile agents computing. However, somewhat surprisingly, there are no results showing a provable difference of cost between the synchronous and asynchronous versions of a task executed by mobile agents. The aim of this paper is to fill up this gap. We show for the first time that for some natural task executed by mobile agents in a network, the optimal cost of its deterministic solution in the asynchronous setting has higher order of magnitude than that in the synchronous scenario. The task for which we show this difference is well-studied: that of rendezvous of two agents in an infinite oriented grid. More precisely, we consider two agents with distinct integer labels starting at a distance D in the infinite oriented grid. Each agent knows D and its own label but not the label of the other agent and it does not know the position of the other agent relative to its own. Agents do not have any global system of coordinates. They have to meet in a node or inside an edge of the grid, and the cost of a rendezvous algorithm is the number of edge traversals by both agents until the meeting. We show that in the synchronous scenario rendezvous can be performed at cost O(D@?), where @? is the length of the (binary representation of the) smaller label, while cost @W(D^2+D@?) is needed for asynchronous completion of rendezvous. Hence, for instances with @?=o(D), the optimal cost of asynchronous rendezvous is asymptotically larger than that of synchronous rendezvous.