Resource Allocation Optimization. Xxxxxxxxx and his team try to optimize database performance in the cloud by partitioning CPU capacity among virtual machines (Xxxxxxxxx et al., 2009). Xxxxxxxxx et al. use white box model and black box model as its cost model to predict the database performance. White box model is based on the internal knowledge of the database system and black box model is statistical model based on external, empirical observations of the database system’s performance. The authors focus on optimizing CPU resource allocation within a physical machine. On the other hand, this thesis not only considers optimization within a physical machine but also multiple physical machines where each containing multiple tenants. The problem of automatically configuring CPU resource allocation to multiple virtual machines that are all running database systems and sharing a pool of physical resources is considered by Soror and his group (Soror et al., 2010). Soror et al. rely on the cost model of the database query optimizers as its cost model to predict the workload performance under different resource allocations. The authors’ work is sensitive to workload resource needs, for example, the system will not allocate more CPU resource to a virtual machine if it does not decrease total execution time. In some cases, the query optimizer cost model is inaccurate, thus affecting the resource allocation decisions. The authors correct the optimizer error through their online refinement approach to fine tune the resource allocation decisions. Kairos, a database consolidation system is presented by Xxxxxx and his team. It uses a technique to measure the hardware requirements of database workloads, as well as models to predict the combined resource utilization of those workloads (Xxxxxx et al., 2011). As database management system is usually operated on a dedicated server, the system is configured to use all available server resources. This caused the amount of memory provisioned far exceeds the actual working set at any point in time. They presented a technique called buffer pool gauging to estimate the exact amount of memory the current database working set is utilizing. Experiment has been conducted to compare the DB-in-VM approach versus the shared process approach, and shows that their approach provides between 6x to 12x higher throughput. Their work also shows that past workload behavior is also a good predictor of its future behavior. Cost-Efficient database Placement (CEP) algorithm is propose...
