Cost-Effective Data Partition for Distributed Stream Processing System

Data skew and dynamics greatly affect throughput of stream processing system. It requires to design a high-efficient partition method to evenly distribute workload in a distributed and parallel. Previous research mainly focuses on load balancing adjustment based on key-as-granularity or tuple-as-granularity, both of which have their own limitations such as clumsy balance activities or expensive network cost. In this paper, we present a comprehensive cost model for partitioning method, which makes a synthesis estimation of memory, CPU and network resource utilization. Based on cost model, we propose a novel load balancing adjustment algorithm, which adopts the idea of “Split keys on demand and Merge keys as far as possible”, and is adaptive to different skewed workload. Our evaluation demonstrates that our method outperforms the state-of-the-art partitioning schemes while maintaining high throughput and resource utilization.