This talk overviews our inquiry in systems software for taming large IoT data. I will describe two systems that target two complementary analytics paradigms: a data engine for processing large telemetry streams (hot data); a data store for querying large archival videos (cold data). The two systems exploit emerging hardware (e.g. 3D-stacked DRAM) as well as emerging workloads (e.g. neural networks). Both systems advance the state-of-the-art performance by orders of magnitude.
Our experiences highlight the significance of designing OSes for specific scenarios, where the OSes play key roles: mapping AI to new hardware, dynamically configuring AI, and trading off among competing objectives.