Obtaining and Managing Answer Quality for Online Data-Intensive Services

Fully autonomous aerial systems (FAAS) fly complex missions guided wholly by software. If users choose software, compute hardware and aircraft well, FAAS can complete missions faster and safer than unmanned aerial systems piloted by humans. On the other hand, poorly managed edge resources slow down missions, waste energy and inflate costs. This paper presents a model-driven approach to manage FAAS. We fly real FAAS missions, profile compute and aircraft resource usage and model expected demands. Naive profiling approaches use traces from previous flights to infer resource usage. However, edge resources can affect where FAAS fly and which data they sense. Usage profiles can diverge greatly across edge management policies. Instead of using traces, we characterize whole flight areas to accurately model resource usage for any flight path. We combine expected resource demands to model mission throughput, i.e., missions completed per fully charged battery. We validated our model by creating FAAS, measuring mission throughput across many system settings. Our FAAS benchmarks, released through our open source FAAS suite SoftwarePilot, execute realistic missions: autonomous photography, search and rescue, and agricultural scouting using well-known software. Our model predicted throughput with 4% error across mission, software and hardware settings. Competing approaches yielded 10-24% error. We used our SoftwarePilot benchmarks to study (1) GPU acceleration, scale up, and scale out, (2) onboard, edge and cloud computing, (3) energy and monetary budgets, and (4) software driven GPU management. We found that model-driven management can boost mission throughput by 10X and reduce costs by 87%.

show abstract

“…• Ubora: Mimicks adaptive quality management in recent research [24,25]. Each mission is treated as a query.…”

Section: Adaptive Hardware-workload Co-designmentioning

confidence: 99%

Managing edge resources for fully autonomous aerial systems

Boubin

Babu

Stewart

et al. 2019

Proceedings of the 4th ACM/IEEE Symposium on Edge Computing

Self Cite

View full text Add to dashboard Cite

show abstract

“…However, delaying updates degrades accuracy. Prior work has shown that staleness corresponds to answer quality [7,9,10]. Hard limits on staleness prevent gross degradation on quality.…”

Section: Motivationmentioning

confidence: 99%

Revisiting online scheduling for AI-driven internet of things

Babu

Stewart

2019

Proceedings of the 4th ACM/IEEE Symposium on Edge Computing

Self Cite

View full text Add to dashboard Cite

AI-driven Internet of Things (IoT) use AI Inference to characterize data processed from various sensors. Together, AI and IoT support smart buildings, cities, cars and drones. However, AI Inference requires updates when executed in new contexts. Frequent updates consume more energy and drain Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for thirdparty components of this work must be honored. For all other uses, contact the owner/author(s).

show abstract

“…However, delaying updates degrades answer quality (i.e., the accuracy of classifications). Staleness corresponds to answer quality [5,6]. Hard limits on staleness can prevent large quality loss.…”

Section: Introductionmentioning

confidence: 99%

Energy, latency and staleness tradeoffs in AI-driven IoT

Babu

Stewart

2019

Proceedings of the 4th ACM/IEEE Symposium on Edge Computing

Self Cite

View full text Add to dashboard Cite

AI-driven Internet of Things (IoT) use AI inference to characterize data harvested from IoT sensors. Together, AI inference and IoT support smart buildings, smart cities and autonomous vehicles. However, AI inference consumes precious energy, drains batteries and shortens IoT lifetimes. Deep sleep modes on IoT processors can save energy during long, uninterrupted idle periods. When AI software is updated frequently, scheduling policies must choose between interrupting deep sleep and degrading AI inference by delaying updates. Scheduling is challenging because of the stochastic nature of update arrivals, processing needs and updates cannot be delayed indefinitely. This paper studies scheduling policies when (1) updates (tasks) arrive frequently, (2) updates must be processed within staleness limits and (3) energy footprint is the metric of merit. We define a scheduling policy as a sequence of choices that decide when updates are applied. We use random walks to explore the space of scheduling policies and 2 K r design of experiments to quantify primary effects and interactions between factors. We conducted 6 2 K r tests with 5X replication each. Each test executes 1,000,000 random walks and computes their energy footprint. We simulated multiple IoT, e.g., varying the number of AI inference components from 5-500. The best random-walk policy uses much less energy than 99 t h and 95 t h percentiles. First-come-first-serve and shortest-job-first policies use 7X more energy than the best policy.

show abstract

Obtaining and Managing Answer Quality for Online Data-Intensive Services

Cited by 12 publications

References 25 publications

Managing edge resources for fully autonomous aerial systems

Managing edge resources for fully autonomous aerial systems

Revisiting online scheduling for AI-driven internet of things

Energy, latency and staleness tradeoffs in AI-driven IoT

Contact Info

Product

Resources

About