| Summary: | 碩士 === 國立清華大學 === 資訊工程學系 === 93 === The data stream management system (DSMS) is an active research area in recent years. Since the number of available resources for a DSMS is limited and the input rates of data streams are unpredictable, ensuring a steady performance of the DSMS when the workload exceeds the system capacity is of critical importance. The previous work, Aurora, well motivates the problem of load shedding under the limited CPU cycles per time unit and proposes an approach to insert some operators for dropping data into the DSMS once the overloading situation is expected to occur. There can be various schemes to insert such operators. In Aurora, all the possible schemes are pre-computed and stored as a lookup table for handling the overloading situation. Due to the variable input rates of data streams, this table needs to be reconstructed. In this thesis, we follow the context of Aurora and propose approaches to adaptively reconstruct that table for efficient load shedding. In our approaches, we devise a novel method to check whether the table for load shedding should be modified by observing the input rates of data streams. However, some schemes in the table to be modified can still be reused. Therefore, we design two approaches to modify only some schemes in the table. One approach produces a table with all possible schemes. The other approach produces a table with only the necessary schemes, which is sufficient to shed the current workload. Experiment results show that both of our approaches outperform the Aurora’s approach on the time for maintaining the table for load shedding.
|
|---|
