Abstract:

The growth of Internet of Things (IoT) deployments drives the need for data management systems that seamlessly integrate real-time sensor streams with historical datasets, without overburdening last-mile networks or incurring prohibitive cloud costs. In this talk, we will first outline the limitations of today’s centralized pipelines, which route every byte of sensor data through cloud datacenters, leading to static resource allocation and escalating network expenses.

Next, we present Adaptive Sensor Data Management (ASDM), our overarching approach that embeds adaptivity at multiple levels of the data path. We will talk about:

• Network-aware trade-offs, where lightweight emulation techniques pinpoint when and how to pre-process or filter data at the edge, preserving query fidelity while slashing bandwidth. (DEBS’20)
• Chameleon, a bidirectional protocol that dynamically tunes sensors in tandem with cloud analytics to maintain low latency and avoid global bottlenecks, enabling up to 80 % more sensors on a fixed network budget with negligible query error. (ICDE’25)
• Near-network offload, which leverages SmartNIC accelerators at ingestion points to push down compression, decompression, and filtering operators—dramatically boosting throughput or reducing CPU load through adaptive co-processing. (VLDB’26)

We will conclude with performance highlights and discuss the open challenges in making ASDM a practical foundation for next-generation data platforms, beyond just IoT scenarios.

Bio:

Dimitrios Giouroukis is a last-year PhD at TU Berlin’s DIMA group. He investigates ways to further scale data management systems even when they operate in last-mile networks. His research interests lie in the intersection of IoT, sensor-cloud query processing, and in-network data management. He mainly works with Varun Pandey, Steffen Zeuch, and Prof. Volker Markl.