SO-1SR:Towards a self-optimizing One-Copy Serializability Protocol for Data Management in the Cloud
Ilir Fetai and Heiko Schuldt
Fifth International Workshop on Cloud Data Management (CloudDB 2013)
Clouds are very attractive environments for deploying different types of applications due to their pay-as-you-go cost model and their highly available and scalable infrastructure. Data management is an integral part of the applications deployed in the Cloud. Thus, it is of utmost importance to provide highly available and scalable data management solutions tailored to the needs of the Cloud. Data availability can be increased by using well-known replication techniques. Data replication also increases scalability in case of read-only transactions, but generates a considerable overhead for keeping the replicas consistent in case of update transactions. In order to meet the scalability demands of their customers, current Cloud providers use DBMSs that only support weak data consistency. While weak consistency is considered to be sufficient for many of the currently deployed applications in the Cloud, more and more applications with strong consistency guarantees, like traditional online stores, are moved to the Cloud. In the presence of replicated data, these applications require one-copy serializability (1SR). Hence, in order to exploit the advantages of the Cloud also for these applications, it is essential to provide scalable, available, low-cost, and strongly consistent data management, which is able to adapt dynamically based on application and system conditions. In this paper, we present SO-1SR (self-optimizing 1SR), a novel customizable load balancing approach to transaction execution on top of replicated data in the Cloud which is able to efficiently use existing resources and to optimize transaction execution in an adaptive and dynamic manner without a dedicated load balancing component. The evaluation of SO-1SR on the basis of the TPC-C benchmark in the AWS Cloud environment has shown that the SO-1SR load balancer is much more efficient compared to existing load balancing techniques.