Improving performance of streaming applications with filtering and control messages

Abstract

In streaming computing applications, some data can be filtered to reduce computation and communication. Due to filtering, however, some necessary information might be lost. To recover lost information, we use control messages, which carry control information rather than input data. The order between control messages and input data must be precise to guarantee correct computations. In this paper, we study the use of control message in suppressing data communication, which improves throughput. To ensure precise synchronization between control messages and input data, we propose a credit-base protocol and prove its correctness and safety. Results show that with the help of control messages, the application throughput can be improved in proportion to filtering ratios.