Toward optimal operator parallelism for stream processing topology with limited buffers

Stream processing is an emerging in-memory computing paradigm to handle massive amounts of real-time data. It is vital to have a mechanism to propose proper parallelism for the operators to handle streaming data efficiently. Previous research has mostly focused on parallelism optimization with infinite buffers; however, the topology’s quality of service is severely affected by network buffers. Thus, in this paper, we introduce an extended queueing network to model the relationship between the parallelism and tuple’s average sojourn time with limited buffers. Based on this model, we also propose greedy algorithms to calculate the optimal parallelism for both the minimum latency and maximum throughput with resource constraints. To fairly evaluate the performance of different models, a random parameter generator for the streaming topology is presented. Experiments show that the extended queuing model may properly forecast performance. Compared to the state-of-the-art method, the proposed algorithms reduce the median total sojourn time by 3.74 times and increase the average maximum sustainable throughput by 1.69 times.