Known Knowns and Unknowns: Near-realtime Earth Observation Via Query Bifurcation in...
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18·
summary
Serval is a system that splits queries into glacial/dynamic filter bifurcation, precomputes stable filters on ground stations and runs only rapidly changing filters on satellites, using Planet Dove traces (10M images, ~200 satellites) and hardware emulations. Evaluation shows 99% compute reduction for many queries and cuts high-priority end-to-end latency from ~71.7 hours to 2 minutes (90th percentile from 149h to 47min) when combined with a global ground-station network (L2D2). It is implemented and validated with trace-driven simulation and emulation.
AI-written summary · about
chapters (13)
- 0:00Near-realtime Earth observation intro
- 2:53Satellite imagery delivery limits
- 6:00Filtering image challenges
- 8:31Traditional approaches fail
- 9:14Servo's image delivery goals
- 10:16Query modeling in Servo
- 12:22Predictability to offload compute
- 13:06Servo system overview
- 13:49Real-world evaluation details
- 14:53Latency improvements with Servo
- 15:37Servo scalability under load
- 15:57Open challenges and future work
- 16:17Conclusion and Q&A
related talks
39mSatellite Imagery to Machine-Readable Intelligence
14mSatellite data monitoring for analytical models
05mAlex Kudriashova Interview
01h 11mProgramming for the planet
44mDemocratizing Climate Science
17mDynamic Scheduling of Approximate Telemetry Queries
10mUbiquitous Satellite Imagery of our Planet
36mRapid Processing of NASA Satellite Data Stored with Ceph