SDPS / PLTC talk by Alberto Lerner

Title

Harnessing Network Power for Data Processing

Speaker

Alberto Lerner is a Maître Assistant at the University of Fribourg in Switzerland. His interests revolve around data-intensive systems and programming models, particularly using heterogeneous hardware devices to support them. He has participated in designing and implementing several programmable systems, both in Academia and companies such as IBM Research, Google, and MongoDB before that. He has also been involved in designing and prototyping computational I/O devices that support data-intensive workloads with better programmability, performance, cost, or energy consumption--sometimes all at once. Alberto has recently co-authored a monograph titled "The Principles of Database and SSDs Codesign" and has been on several Program Committees for the Database and Systems communities, including SIGMOD, VLDB, CIDR, EDBT, ICDE, and Usenix ATC, receiving distinguished reviewer awards in some of them. He has also won a best reproducibility paper award in VLDB 2020.

Abstract

A network can be seen as a powerful computer that connects other computers. It is, however, oddly dimensioned since it has tremendous I/O capacity but relatively simple processing capabilities. This design was meant to allow networking protocols to be implemented. In this talk, we show that the network can be much more. By slightly reprogramming some networking devices, the network can become a tremendously strategic site for data processing. We will illustrate this through two use cases. In the first case, we show how a slightly modified network switch can perform aggregations and all-reduce operations essential in, for instance, analytical databases and distributed machine learning training. In the second case, we discuss how an edge network interface can perform data cleaning on complex sensor data such as LiDAR point clouds, allowing the data to be used later, also in training. If time permits, we will demystify a technology called FPGA that can be used to prototype these types of devices and that, thanks to its declarative programming approach, may sound familiar to systems practitioners who work close to hardware.

Host

Professor Martin Elsman.