Efficient P2P Inspired Policy to Distribute Resource Information in Large Distributed Systems

P. VERGHELET; E. MOCSKOS

High Performance Computing

Springer International Publishing

Lugar: Cham; Año: 2017; p. 3 - 17

The computational infrastructures are becoming larger and more complex. Their organization and interconnection are acquiring newdimensions with the increasing adoption of Cloud Technology and thestart of establishment of Federations of cloud providers.These large interconnected systems require monitoring at different levelsof the infrastructure: from the availability of hardware resources to theeffective provision of services and verification of terms of the establishedagreements.Monitoring becomes a fundamental component of any Cloud Service orFederation, as the up-to-date information about resources in the systemis extremely important to be used as an input to the scheduler component. The way in which the different members of such a distributedsystem obtains and distributes the resource information is what is knownas Resource Information Distribution Policy.Moving towards the obtantion of a scalable and easy to maintain policyleads to the interaction with the Peer to Peer (P2P) paradigm. Someof the proposed policies are based on establishing a ranking accordingto previous communications between nodes. These policies are known aslearning based methods or Best-Neighbor (BN). However, the use of thistype of policies shows poor performance and limited scalability  compared with defacto Hierarchical or other hybrid policies.In this work, we introduce pBN which is a fully distributed resourceinformation policy based on P2P. We analyze some reasons that couldproduce the poor performance in standard BN and propose an improve-ment which shows performance and bandwidth consumption similar toHierarchical policy and other hybrid variations. To compare the differentpolicies, a specific simulation tool is used with different system sizes andexponential network topology.