Adaptive Partitioning and Order-Preserved Merging of Data Streams

Partitioning is a key concept for utilizing modern hardware, especially to exploit parallelism opportunities from many-core CPUs. In data streaming scenarios where parameters like tuple arrival rates can vary, adaptive strategies for partitioning solve the problem of overestimating or underestimating query workloads. While there are many possibilities to partition the data flow, threads running partitions independently from each other lead to unordered output inevitably. This is a considerable difficulty for applications where tuple order matters, like in stream reasoning or complex event processing scenarios. In this paper, we address this problem by combining an adaptive partitioning approach with an order-preserving merge algorithm. Since reordering output tuples can only worsen latency, we mainly focus on the throughput of queries while keeping the delay on individual tuples minimal. We run micro-benchmarks as well as the Linear Road benchmark, demonstrating correctness and effectiveness of our approach while scaling out on a single Xeon Phi many-core CPU up to 256 partitions.