20170222-Kamal-Thesis.pdf (3.97 MB)
Download file

Workload-Aware Data Management in Shared-Nothing Distributed OLTP Databases

Download (7.92 MB)
thesis
posted on 02.03.2017, 22:27 by Joarder Mohammad Mustafa Kamal
More than 3.26 billion Internet users are using Web and mobile applications every day for retail services and businesses processing; information management and retrievals; media and social interactions; gaming and entertainments, etc. With the rapid growth in simultaneous users and data volume, such applications often rely on distributed databases to handle large magnitudes of On-Line Transaction Processing (OLTP) requests at the Internet scale. These databases employ user-defined partitioning schemes during their initial application design phase to scale-out concurrent writing of user data into a number of shared-nothing servers. However, processing Distributed Transactions (DTs), that involve data tuples from multiple servers, can severely degrade the performance of these databases. In addition, transactional patterns are often influenced by the frequent data tuples in the database, resulting in significant changes in the workload that none of the static partitioning schemes can handle in real-time. This thesis addresses the abovementioned challenge through workload-aware incremental data repartitioning. The primary goal is to minimally redistribute data tuples within the system with a view to minimising the adverse impact of DT without adversely affecting the data distribution load balance. Furthermore, to support both range and consistent-hash based data partitions used in OLTP systems, a distributed yet scalable data lookup process, inspired by the roaming protocol in mobile networks, is introduced. Moreover, additional challenge lies on how to reduce the computational complexities while processing large volume of transactional workload very quickly to make real-time decisions. These complexities are addressed in three distinct ways — hide the workload network complexities during analysis by using different level of abstractions; find the appropriate size of observation window to decide how far to look back in the past for workload analysis; and finally, incrementally (i.e., how frequently) trigger the repartitioning process based on a threshold. Some well-defined Key Performance Indicator (KPI) metrics for database repartitioning are defined and rigorous sensitivity analyses are performed on different types of transaction generation models and observation window sizes. In the first stage, graph-theoretic higher level abstractions such as graphs, hypergraphs, and their compressed forms are used to deliver sub-optimal repartitioning decisions using graph min-cut techniques. In order to deliver optimal decisions, a greedy approach is further developed, which ranks the transactions in terms of the measure of repartitioning objectivity and performs transaction-level optimisation to maximise a particular KPI. Finally, transactional stream mining is used to construct a representative sub-network of frequent transactions and perform transaction-level adaptive repartitioning, which eventually reduces the computational complexity, while concomitantly improving the performance due to filtering out a significant volume of infrequent transactions, that have little bearing on the KPIs. Simulation-based experimental evaluations demonstrate that the proposed workload-aware incremental repartitioning schemes outperform the existing static and dynamic repartitioning approaches to be considered effective for production deployment.

History

Campus location

Australia

Principal supervisor

Manzur Murshed

Additional supervisor 1

Joarder Kamruzzaman

Additional supervisor 2

Rajkumar Buyya

Additional supervisor 3

Mohamed Gaber

Year of Award

2017

Department, School or Centre

Gippsland School of Information Technology

Faculty

Faculty of Information Technology