COLA: Optimizing Stream Processing Applications Via Graph Partitioning

In this paper, we describe an optimization scheme for fusing compile-time operators into reasonably-sized run-time software units called processing elements (PEs). Such PEs are the basic deployable units in System S, a highly scalable distributed stream processing middleware system. Finding a high quality fusion significantly benefits the performance of streaming jobs. In order to maximize throughput, our solution approach attempts to minimize the processing cost associated with inter-PE stream traffic while simultaneously balancing load across the processing hosts. Our algorithm computes a hierarchical partitioning of the operator graph based on a minimum-ratio cut subroutine. We also incorporate several fusion constraints in order to support real-world System S jobs. We experimentally compare our algorithm with several other reasonable alternative schemes, highlighting the eectiveness of our approach.