Adaptability Experiments in the RAID Distributed Database System

Adaptable systems can improve reliability and performance by allowing dynamic reconfiguration. We are conducting a series of experiments on the RAID distributeel database system to siudy the cost and performance impHcations of providing static and dynamic adaptability, and for increasing the availabiliiy of daia iiems. Our studies of the cost of our adaptable implementation wefe conducted in the context of the concurrency controller and the replication controller. The experimentation with dynamic adaptability focuses on concurrency conlrol, and our examination of the costs of providing greater data availability studies thf! replication control and atomicity control subsystems of RAID. We show that for concurrency control and replicatioll control, adaptable implementations can be provided at costs comparable to those of speeial purpose algorithms. We also show that for our concurrency controller dynamic adaptability can result in performance benefits and that system reconfiguration can be accomplished dynamically with less cost than stopping the system, performing reconfiguration, and then restarting the syslem. In some cases, reconfiguration could be performed without aborting any transactions. We demonstrate some costs associated with increasing availability through replication control methods and use of a three-phase commil protocol. A system that can dynamically change to algorithms that increase availability can result in a 25-50% performance improvement over systems that continuously employ the algorithms that provide tIle better availallility. We show that the algorithms selected for replication control can significantly impact the time required fOf transaction commilmeul.