Performance Implications of Using Diverse Redundancy for Database Replication

Using diverse redundancy for database replication is the focus of this thesis. Traditionally, database replication solutions have been built on the fail-stop failure assumption, i.e. that crashes are believed to cause a majority of failures. However, recent findings refuted this common assumption, showing that many of the faults cause systematic non-crash failures. These findings demonstrate that the existing, non-diverse database replication solutions, which use the same database server products, are ineffective fault-tolerant mechanisms. At the same time, the findings motivated the use of diverse redundancy (when different database server products are used) as a promising way of improving dependability. It seems that using a faulttolerant server, built with diverse database servers, would deliver improvements in availability and failure rates compared with the individual database servers or their replicated, non-diverse configurations. Besides the potential for improving dependability, one would like to evaluate the performance implications of using diverse redundancy in the context of database replication. This is the focal point of the research. The work performed to that end can be summarised as follows: We conducted a substantial performance evaluation of database replication using diverse redundancy. We compared its performance to the ones of various nondiverse configurations as well as non-replicated databases. The experiments revealed systematic differences in behaviour of diverse servers. They point to the potential for performance improvement when diverse servers are used. Under particular workloads diverse servers performed better than both non-diverse and non-replicated configurations. We devised a middleware-based database replication protocol, which provides dependability assurance and guarantees database consistency. It uses an eager update everywhere approach for replica control. Although we focus on the use of diverse database servers, the protocol can be used with the database servers from the same vendor too. We provide the correctness criteria of the protocol. Different regimes of operation of the protocol are defined, which would allow it to be dynamically optimised for either dependability or performance improvements. Additionally, it can be used in conjunction with high-performance replication solutions. We developed an experimental test harness for performance evaluation of different database replication solutions. It enabled us to evaluate the performance of the diverse database replication protocol, e.g. by comparing it against known replication solutions. We show that, as expected, the improved dependability exhibited by our replication protocol carries a performance overhead. Nevertheless, when optimised for performance improvement our protocol shows good performance. In order to minimise the performance penalty introduced by the replication we propose a scheme whereby the database server processes are prioritised to deliver performance improvements in cases of low to modest resource utilisation by the database servers. We performed an uncertainty-explicit assessment of database server products. Using an integrated approach, where both performance and reliability are considered, we rank different database server products to aid selection of the components for the fault-tolerant server built out of diverse databases. Chapter