Fine-Grained Multi-Query Stream Processing on Integrated Architectures
Fine-Grained Multi-Query Stream Processing on Integrated Architectures
复制标题
集成架构上的细粒度多查询流处理
DOI：
10.1109/tpds.2021.3066407
复制
发表时间：
2021
影响因子：
5.3
通讯作者：
Du Xiaoyong
中科院分区：
文献类型：
--
作者：
Zhang Feng;Zhang Chenyang;Yang Lin;Zhang Shuhao;He Bingsheng;Lu Wei;Du Xiaoyong
探索多个流查询之间的共享机会对于高性能流处理至关重要。现代流处理需要通过使用异构协处理器（如GPU）来加速多个查询，这已被证明是一种有效的方法。新兴的CPU-GPU集成架构6将CPU和GPU集成在同一芯片上，消除了PCI-e带宽瓶颈。这种新颖的体系结构提供了新的机会，提高多查询性能的流处理，但还没有充分探讨现有的系统。我们介绍了一个流处理引擎，称为FineStream，高效的多查询窗口为基础的流处理的CPU-GPU集成架构。FineStream的主要贡献是在CPU和GPU之间提供了一种新的细粒度工作负载调度机制，以利用这两种架构。特别是，FineStream能够有效地处理静态和动态流中的多个查询。实验结果表明：1）在集成架构下，FineStream的吞吐量平均提高了52%，延迟降低了36%; 2）与采用不同设备进行多个查询的粗粒度策略相比，FineStream的吞吐量提高了32%; 3）与分立架构上的流处理引擎相比，集成架构上的FineStream实现了10.4倍的价格吞吐率，1.8倍的能效，并可以享受更低的延迟优势。
Exploring the sharing opportunities among multiple stream queries is crucial for high-performance stream processing. Modern stream processing necessitates accelerating multiple queries by utilizing heterogeneous coprocessors, such as GPUs, and this has shown to be an effective method. Emerging CPU-GPU integrated architectures 6integrate CPU and GPU on the same chip and eliminate PCI-e bandwidth bottleneck. Such a novel architecture provides new opportunities for improving multi-query performance in stream processing but has not been fully explored by existing systems. We introduce a stream processing engine, called FineStream, for efficient multi-query window-based stream processing on CPU-GPU integrated architectures. FineStream's key contribution is a novel fine-grained workload scheduling mechanism between CPU and GPU to take advantage of both architectures. Particularly, FineStream is able to efficiently handle multiple queries in both static and dynamic streams. Our experimental results show that 1) on integrated architectures, FineStream achieves an average 52 percent throughput improvement and 36 percent lower latency over the state-of-the-art stream processing engine; 2) compared to the coarse-grained strategy of applying different devices for multiple queries, FineStream achieves 32 percent throughput improvement; 3) compared to the stream processing engine on the discrete architecture, FineStream on the integrated architecture achieves 10.4× price-throughput ratio, 1.8× energy efficiency, and can enjoy lower latency benefits.