Control-Based Database Tuning Under Dynamic Workloads

Database administration (tuning) is the process of adjusting database configurations in order to accomplish desirable performance goals. This job is performed by human operators called database administrators (DBAs) who are generally well-paid, and are becoming more and more expensive with the increasing complexity and scale of modern databases. There has been considerable effort dedicated to reducing such cost (which often dominates the total ownership cost of missioncritical databases) by making database tuning more automated and transparent to users (Chaudhuri et al, 2004; Chaudhuri and Weikum, 2006). Research in this area seeks ways to automate the hardware deployment, physical database design, parameter configuration, and resource management in such systems. The goal is to achieve acceptable performance on the whole system level without (or with limited) human intervention. According to Weikum et al. (2002), problems in this category can be stated as: workload × configuration (?) ? performance which means that, given the features of the incoming workload to the database, we are to find the right settings for all system knobs such that the performance goals are satisfied The following two are representatives of a series of such tuning problems in different databases: • Problem 1: Maintenance of multi-class servicelevel agreements (SLA) in relational databases. Database service providers usually offer various levels of performance guarantees to requests from different groups of customers. Fulfillment of such guarantees (SLAs) is accomplished by allocating different amounts of system resources to different queries. For example, query response time is negatively related to the amount of memory buffer assigned to that query. We need to dynamically allocate memory to individual queries such that the absolute or relative response times of queries from different users are satisfied. • Problem 2: Load shedding in stream databases. Stream databases are used for processing data generated continuously from sources such as a sensor network. In streaming databases, data processing delay, i.e., the time consumed to process a data point, is the most critical performance metric (Tatbul et al., 2003). The ability to remain within a desired level of delay is significantly hampered under situations of overloading (caused by bursty data arrivals and time-varying unit data processing cost). When overloaded, some data is discarded (i.e., load shedding) in order to keep pace with the incoming load. The system needs to continuously adjust the amount of data to be discarded such that 1) delay is maintained under a desirable level; 2) data is not discarded unnecessarily.