Data Stream Processing engines have been recently proposed as a powerful tool to facilitate the analysis of network telemetry data. Motivated by the increasing amount of data to be analyzed, state-of-the-art approaches couple them with programmable switches to filter-out uninteresting traffic and thus helping scaling-out their processing capabilities. In this paper, we propose the use of SmartNICs as efficient accelerators of stream processing operators, instead. SmartNICs are commonly deployed in datacenter networks and their architecture, composed by many low power-processors, well aligns with the highly-parallelizable computational processing required by standard frameworks for streaming analysis. We started from WindFlow, a state-of-the-art stream processor, and developed a flow meter monitoring application on top of it. We offloaded part of its computation to a commodity Netronome and to demonstrate the generality of our approach, we implemented our offload in eBPF so that our logic can be ported to any NIC supporting this programming paradigm. We show that our solution can analyze 1.6× more traffic than a pure software approach.
SmartNIC-Accelerated Stream Processing Analytics
Lettieri, Giuseppe;Fais, Alessandra;Procissi, Gregorio
2023-01-01
Abstract
Data Stream Processing engines have been recently proposed as a powerful tool to facilitate the analysis of network telemetry data. Motivated by the increasing amount of data to be analyzed, state-of-the-art approaches couple them with programmable switches to filter-out uninteresting traffic and thus helping scaling-out their processing capabilities. In this paper, we propose the use of SmartNICs as efficient accelerators of stream processing operators, instead. SmartNICs are commonly deployed in datacenter networks and their architecture, composed by many low power-processors, well aligns with the highly-parallelizable computational processing required by standard frameworks for streaming analysis. We started from WindFlow, a state-of-the-art stream processor, and developed a flow meter monitoring application on top of it. We offloaded part of its computation to a commodity Netronome and to demonstrate the generality of our approach, we implemented our offload in eBPF so that our logic can be ported to any NIC supporting this programming paradigm. We show that our solution can analyze 1.6× more traffic than a pure software approach.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.