An simulation approach for evaluating scalability of a Virtually Fully Replicated real-time database

To achieve timeliness for transactions in a distributed real-time main memory database, we store each database replica at every node, allowing all transactions to be executed locally. With detached replication, updates are not replicated to other nodes within the execution of the transaction, but are logged and replicated at low priority when system resources allow that. In such a system concurrent updates of different data object replicas introduces inconsistencies, which can be resolved by a conflict detection and resolution mechanism [9]. We have chosen to study scalability for replicated databases with detached replication. Full replication of the database uses far more system resources than actually is needed for providing applications at nodes with the data objects used by their transactions. With Virtual Full Replication only the data objects that are used by transactions at each node are replicated. Typically this is a only a small subset of the entire database. Virtual Full Replication provides the image to the application of a fully replicated database, maintaining the advantages of full replication while reducing the resource requirements, and thereby improves scalability. We have implemented Virtual Full Replication in the DeeDS database prototype [1] by segmentation of the database, based the specification of the data accesses and other data requirements that transactions have [16]. With this implementation we evaluated the usage of three key resources: bandwidth, storage and processing. It was found that with Virtual Full Replication resource usage grows with the bound of the replication degree rather than with the square of typical system parameters, such as the number of nodes, the frequency of transactions or the size of the database. In scalable systems, the growth of the usage of system resources can not be larger than the change in the system parameters. Otherwise the system grows out of resources and is not scalable. Our current implementation can only be run with few nodes due to limited amount of resources available, in terms of number of nodes and the amount of available memory. Thus a large scale system can not be evaluated. By representing the resource usage by variables in a simulation, rather than actually using the resources, we may be able to examine very large systems. This paper describes our approach for simulating an implementation of static segmentation in a distributed real-time main memory database. Our approach consist both of a literature review of validation in simulations, and also how such simulation approach could be done for the system we study. Currently we have a part implementation of replication mechanisms required in the system to be modeled, and therefore we cannot yet state how useful such a simulation is.