Skip to Main content Skip to Navigation
Theses

Data replication in large-scale data management systems

Abstract : In recent years, growing popularity of large-scale applications, e.g. scientific experiments, Internet of things and social networking, led to generation of large volumes of data. The management of this data presents a significant challenge as the data is heterogeneous and distributed on a large scale. In traditional systems including distributed and parallel systems, peer-to-peer systems and grid systems, meeting objectives such as achieving acceptable performance while ensuring good availability of data are major challenges for service providers, especially when the data is distributed around the world. In this context, data replication, as a well-known technique, allows: (i) increased data availability, (ii) reduced data access costs, and (iii) improved fault-tolerance. However, replicating data on all nodes is an unrealistic solution as it generates significant bandwidth consumption in addition to exhausting limited storage space. Defining good replication strategies is a solution to these problems. The data replication strategies that have been proposed for the traditional systems mentioned above are intended to improve performance for the user. They are difficult to adapt to cloud systems. Indeed, cloud providers aim to generate a profit in addition to meeting tenant requirements. Meeting the performance expectations of the tenants without sacrificing the provider's profit, as well as managing resource elasticities with a pay-as-you-go pricing model, are the fundamentals of cloud systems. In this thesis, we propose a data replication strategy that satisfies the requirements of the tenant, such as performance, while guaranteeing the economic profit of the provider. Based on a cost model, we estimate the response time required to execute a distributed database query. Data replication is only considered if, for any query, the estimated response time exceeds a threshold previously set in the contract between the provider and the tenant. Then, the planned replication must also be economically beneficial to the provider. In this context, we propose an economic model that takes into account both the expenditures and the revenues of the provider during the execution of any particular database query. Once the data replication is decided to go through, a heuristic placement approach is used to find the placement for new replicas in order to reduce the access time. In addition, a dynamic adjustment of the number of replicas is adopted to allow elastic management of resources. Proposed strategy is validated in an experimental evaluation carried out in a simulation environment. Compared with another data replication strategy proposed in the cloud systems, the analysis of the obtained results shows that the two compared strategies respond to the performance objective for the tenant. Nevertheless, a replica of data is created, with our strategy, only if this replication is profitable for the provider.
Document type :
Theses
Complete list of metadatas

Cited literature [100 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-01820748
Contributor : Abes Star :  Contact
Submitted on : Friday, June 22, 2018 - 9:49:05 AM
Last modification on : Sunday, August 16, 2020 - 3:31:08 AM
Long-term archiving on: : Tuesday, September 25, 2018 - 8:52:36 AM

File

2017TOU30066B.pdf
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-01820748, version 1

Collections

Citation

Uras Tos. Data replication in large-scale data management systems. Web. Université Paul Sabatier - Toulouse III, 2017. English. ⟨NNT : 2017TOU30066⟩. ⟨tel-01820748⟩

Share

Metrics

Record views

947

Files downloads

1677