Many emerging applications analyze data streams by running graphs of communicating tasks called operators. To develop and deploy such applications, Stream Processing Systems (SPSs) like Apache Storm and Flink have been made available to researchers and practitioners. They exhibit imperative or declarative programming interfaces to develop operators running arbitrary algorithms working on structured or unstructured data streams. In this context, the interest in leveraging hardware acceleration with GPUs has become more pronounced in high-throughput use cases. Unfortunately, GPU acceleration has been studied for relational operators working on structured streams only, while non-relational operators have often been overlooked. This paper presents WindFlow, a library supporting the seamless GPU offloading of general partitioned-stateful operators, extending the range of operators that benefit from hardware acceleration. Its design provides high throughput still exposing a high-level API to users compared with the raw utilization of GPUs in Apache Flink.

Modern multi-core servers are powerful enough to process multi-gigabit live packet streams on the network data plane. However, in most cases network programmers must build their applications from scratch, by implementing both the interfaces towards the lower hardware level and the proper mechanisms for parallel programming. Data Stream Processing (DaSP) frameworks have recently emerged as promising approaches to overcome the above issues and to let programmers simply focus on the logic of the application to develop. However, DaSP platforms are generally not designed for the networking domain, in terms of both performance and functions. In this paper, we selected the WindFlow DaSP framework and built suitable extensions to attach multiple (accelerated) packet sources of data to it. We then implemented a simple monitoring application on top of WindFlow and carried out stress tests with synthetic and real traffic. The results prove that performance scale linearly with the processing cores so that the application was able to process the whole amount of live data up to nearly 20 Gbps rate.

Nowadays, we are witnessing the diffusion of Stream Processing Systems (SPSs) able to analyze data streams in near realtime. Traditional SPSs likeStormandFlinktarget distributed clusters and adopt thecontinuous streaming model, where inputs are processed as soon as they are available while outputs are continuously emitted. Recently, there has been a great focus on SPSs for scale-up machines. Some of them (e.g.,BriskStream) still use the continuous model to achieve low latency. Others optimize throughput with batching approaches that are, however, often inadequate to minimize latency for live-streaming applications. Our contribution is to show a novel software engineering approach to design the runtime system of SPSs targeting multicores, with the aim of providing a uniform solution able to optimize throughput and latency. The approach has a formal nature based on the assembly of components calledbuilding blocks, whose composition allows optimizations to be easily expressed in a compositional manner. We use this methodology to build a new SPS calledWindFlow. Our evaluation showcases the benefits ofWindFlow: it provides lower latency than SPSs for continuous streaming, and can be configured to optimize throughput, to perform similarly and even better than batch-based scale-up SPSs.