Coordinated Load Sharing for Large Distributed Systems

  • Ali Abdullah Mohammad Al-Raqbani

Student thesis: Master's Thesis

Abstract

Ali Abdullah lVlohammad Al-Raqbani. Coordinated Load Sharing for Large Distributed Systems. M.Sc. By Research Thesis, Computer Engineering Department, Etisalat University College, Sharjah, United Arab Emirates, December 2005 Advances in high speed computer netvvorks and the availability of low cost powerful pes and workstations have spurred the movement of both business and academic high performance computing from a traditional centralized main frame environment to a decentralized distributed processing environment. By sharing system resources on loosely coupled processors, the computing cost per user can be significantly decreased. However, in order to achieve and support a high level of resource sharing and provide adequate user response time, efficient techniques in distributed load sharing become a requirement. Distributed load sharing can simply be described as an algorithm where jobs or tasks are transparently migrated and subsequently executed at neighbouring processors with intention to ma'{imize global system throughput and meet its response time requirements. In this thesis a new coordinated load sharing scheme is proposed and its performance is analyzed. The scheme is based on decomposing the system into partitions. Each partition contains nodes that are physically close to each other in order to minimize communication overhead resulting from load sharing activities. Each partition has a coordinator to control load sharing inside the partition as well as betvveen partitions. In addition, the coordinator's job is to collect and disseminate load state information among the partition nodes. Unlike the existing systems, the proposed scheme is semicentralized in the sense that load state information is centrally gathered at the coordinator and periodically propagated to all nodes to allow a decentralized load sharing decision. This unique feature of the system enables benefits of both centralized and decentralized load sharing scheme, while it reduces significantly communication overhead and avoids bottlenecks. At the partition level, the same concept is applied where a global coordinator is used to control the load sharing between partitions and to collect the load state of all the partitions in the system. These are periodically broadcasted to all coordinators in the system. The proposed algorithm has been designed to be simple and consistent by adopting a single intra/inter-partition load sharing strategy. The algorithms based on the proposed scheme are simulated and analyzed using the O:NfNeT++ simulator. The performance analysis of the algorithm is done in terms of design of various parameters. They are also compared to the probing based load sharing algorithms. The results have shown that the coordinated load sharing scheme outperforms the probing-based load sharing scheme for highly loaded systems. The suitability of the algorithms for the large distributed system is analyzed and tested under multiple LANs and the results prove that, the proposed scheme is suitable for large distributed system.
Date of Award2005
Original languageAmerican English
SupervisorHassan Barada (Supervisor)

Keywords

  • Distributed Systems
  • Load sharing
  • Load Balancing
  • Distributed Scheduling

Cite this

'