A practical evaluation of load shedding in data stream management systems for network monitoring

In network monitoring, an important issue is the number of tuples the data stream management system (DSMS) can handle for different network loads. In order to gracefully handle overload situations, some DSMSs are equipped with a tuple dropping functionality, also known as load shedding. These DSMSs register and relate the number of received and dropped tuples, i.e., the relative throughput, and perform different kinds of calculations on them. Over the past few years, several solutions and methods have been suggested to efficiently perform load shedding. The simplestapproach is to keep a count of all the dropped tuples, and to report this to the end user. In our experiments, we study two DSMSs, i.e., TelegraphCQ with support for load shedding, and STREAM without this support. We use three particular network monitoring tasks to evaluate the two DSMS with respect to their ability of load shedding and performance. We demonstrate that it is impor- tant to investigate the correctness of load shedding by showing that the reported number of dropped tuples is not always correct.