Accelerating real-time data-analytics with in-network computing
Real-time data-analytics builds on distributed stream processing to provide critical insights for enabling automation and strategic decision-making over continuously incoming data. With the substantial increase of the amount of information and the emergence of time-sensitive operations, existing stream processing systems struggle to fulfil stringent performance requirements such as low-latency, high throughput and state-consistency. This is due to the limitations of host-centric processing that depends on Central Processing Units (CPUs) involving several data copies and unpredictable delays. This research shows that the use of specialized network hardware, especially programmable packet processors, represents a promising alternative for improving the performance of real-time data-analytics.
In this thesis, we investigate a new paradigm called in-network computing with the goal to understand its potential for enhancing the efficiency of distributed stream processing. It proposes the hardware acceleration of data-analytics functions using in-network programming model P4. First, QoS-P4bSub is introduced to enforce strict Quality of Service (QoS) guarantees for delay-sensitive Publish/Subscribe communications. Second, S4 (Stateful and Scalable Splitter Switch) is proposed to enable massive data parallelism and increase the overall system throughput. Lastly, P4mCast (P4-based Atomic Multicast) is presented for supporting total order message delivery for state-consistency.
The performance analysis showed that QoS-P4bSub provides better latency determinism compared to state-of-the-art solution, with 64% reduction in the tail latency. Moreover, S4 supports up to 286k concurrent data streams with a parallelism degree of up to 457k operator instances, and P4mCast handles up to 143k concurrent messages, achieving a throughput of up to 52 billion messages per second.