On supporting the quality of real-time data services.

摘要

Real-time data services can significantly benefit data-intensive real-time applications, such as e-commerce, transportation control, and target tracking, via timely transaction processing using fresh data such as the current stock prices and traffic status data. However, it is challenging to process transactions in a timely fashion using fresh data, since database workloads may considerably vary from time to time. Further, transaction timeliness and data freshness requirements may compete for system resources. To address this problem, we propose to develop (1) novel control theoretic models to capture real-time data service workload characteristics, (2) develop feedback-based QoS (Quality of Service) management schemes to adapt the system behavior under overload, if necessary, to support the desired data service delay, and (3) thoroughly evaluate the performance in a realistic testbed. To this end, we have developed several feedback control models and QoS management schemes specific to real-time data services. Furthermore, we have developed a real-time data service testbed, in which clients submit thousands of data service requests per second to our QoS-aware database developed atop BerkeleyDB---an open source database from Oracle. In the experiments performed in our testbed, the proposed approaches can closely support the desired response time bound, while substantially increasing the database throughput compared to the tested baselines representing the current state of art.