Coordination mechanisms for decentralized parallel systems

Autor(es)

Cohen, Johanne ^[1] ; Cordeiro, Daniel ^[2] ; Trystram, Denis ^{[3, 4]}

Afiliação do(s) autor(es)

^[1] Univ Paris 11, LRI, UMR 8623, F-91400 Orsay - France

^[2] Univ Sao Paulo, Dept Comp Sci, BR-05508090 Sao Paulo - Brazil

^[3] Inst Univ France, F-75005 Paris - France

^[4] Grenoble Univ, LIG, F-38330 Montbonnot St Martin - France

• Tipo de documento: Artigo Científico
• Fonte: CONCURRENCY AND COMPUTATION-PRACTICE & EXPERIENCE; v. 27, n. 5, p. 1255-1272, APR 10 2015.
• Citações Web of Science: 1

Resumo

On resource sharing platforms, the execution of the jobs submitted by users is usually controlled by a centralized global scheduler. It determines efficient schedules regarding some common objective function that all organizations agree with (for instance, maximizing the utilization of the entire platform). However, in practice, each organization is mostly interested in the performance obtained for its own jobs. We study the price that the collectivity must pay in order to allow independence to selfish, self-governing organizations, so they can choose the best schedules for their own jobs. In other words, we are interested in analyzing the costs on the global performance inflicted by the decentralization of scheduling policies. We present a game-theoretic model for the problem and the associated coordination mechanisms developed to reduce the cost of the decentralization of the decision-making process. The main contribution is to show (in theory and practice) how to devise pure Nash equilibria configurations for every instance of the problem and to prove that the price paid by the collectivity depends on the local scheduling policy and on the characteristics of the workload executed on such platforms. Copyright (C) 2014 John Wiley \& Sons, Ltd. (AU)

• Processo FAPESP: 12/03778-0 - Aplicações de Teoria de Escalonamento em Computação em Nuvem
• Beneficiário: Daniel de Angelis Cordeiro
• Modalidade de apoio: Bolsas no Brasil - Pós-Doutorado