Abstract

Many real-time applications (e.g. financial data analytics, online machine learning, and system monitoring) need to process large amounts of streaming data with high throughput and low latency. Window queries (for example, sliding-window aggregates) are fundamental for data stream processing. However, despite the computational power of modern data center servers, nowadays equipped with heterogeneous processors such as CPUs and GPGP Us, window-based query applications remain compute-bound as they lack support to harness these available resources.

This talk introduces Saber, a hybrid data-parallel stream processing system implemented in Java that is capable of using both a multi-core CPU and a GPGPU to accelerate window queries. Rather than pinning (part of) a query on a particular processor, Saber runs it on both and employs a novel preference-based scheduling mechanism to maximise the share of computation that runs on the best-suited processor. When running on the GPGPU , Saber interlaces the execution of consecutive data-parallel tasks to consume the excess system bandwidth that would otherwise remain unused due to the latency of moving data from Java memory to the accelerator (and vice versa). Putting all together, Saber can process typical window queries at line-rate, shifting the erstwhile computational bottleneck to the network.

Bio: Alexandros Koliousis is a Research Associate at the Department of Computing, Imperial College London. He is a member of the Experimental Software Systems and Large-Scale Distributed Systems groups. His research is currently focused on engineering data parallel processing systems that make best use of heterogeneous processors. Prior to joining Imperial College, he was working on home network management at the School of Computing Science, University of Glasgow, from where he also obtained his MSc and PhD degrees.