Odessa Enabling Interactive Perception Applications on Mobile Devices
- 1. Odessa Enabling Interactive Perception Applications on Mobile Devices Miroslav Cupák Moo-Ryong Ra, Anmol Sheth, Lily Mummertx, Padmanabhan Pillai, David Wetherallo, Ramesh Govindan 25/03/2013 Some slides borrowed from http://catnip.usc.edu/home/.
- 2. Outline ● Problem & Contributions ● Metrics ● Applications ● Sprout ● Application Performance ● Design & Implementation ● Evaluation ● Conclusions
- 3. Problem ● resource constrained mobile devices ● offloading and parallelism ● interactive perception applications ● requirements ○ crisp response ○ continuous processing ○ compute-intensive ○ performance depending on data
- 4. Contributions ● study of offloading & parallelism decisions ● Odessa ○ runtime automatically determining how best to use offloading and parallelism to achieve responsiveness and accuracy ● successful evaluation of the greedy approach
- 5. Metrics ● makespan ○ time to execute all stages of computation for a frame ● throughput ○ rate at which frames are processed ● techniques ○ code offloading ○ pipelining ○ data parallelism
- 6. Applications
- 7. Sprout ● distributed stream processing system ● uses a data flow model to dynamically distribute processing of an application ● takes care of data transfer, coarse-grained data parallelism, pipeline parallelism ● Odessa ○ what/when/how to offload on top of Sprout
- 8. Application Performance ● observations ○ offloading must be made dynamically (input variability, network bandwidth, device parameters) ○ data parallelism cannot be determined apriori ○ static choice of pipeline parallelism is not optimal ● setup ○ netbook (1 core), laptop (2 cores), server (8 cores) ○ 30 fps, 640x480 px video, 3000 (500) frames ○ network bandwidth emulation
- 10. Variability Across Mobile Platforms
- 12. Data & Pipeline Parallelism
- 13. Design ● goals ○ minimize makespan and maximize throughput ○ quick response to environment changes ○ low computation and communication overhead ● parts ○ application profiler ○ decision engine
- 14. Profiler ● lightweight, online ● piggybacking approach ● function ○ maintain application performance profile for the decision engine ● data ○ execution time of each processing stage ○ wait time on connectors ○ transfer time
- 15. Decision Engine ● runs periodically ● greedy incremental approach ● estimates impact of offloading and parallelism based on processor frequencies and history of network measurements ● dynamic pipelining based on tokens
- 16. Decision Engine
- 17. Evaluation: Performance & Overhead ● minimum overhead
- 18. Evaluation: Other Strategies ● 3 strategies : offload all, domain-specific, offline optimizer
- 19. Evaluation: Context Adaptation ● taking fidelity into account left for future work
- 20. Related Work ● energy saving: MAUI ● graph-based partitioning: Coign ● static partitioning: Wishbone ● a set of pre-specified partitions: CloneCloud, Chroma, Spectra ● parallel processing: MapReduce
- 21. Conclusions ● understanding of the factors which contribute to offloading and parallelism decisions ● Odessa ○ runtime automatically determining how best to use offloading and parallelism to achieve responsiveness and accuracy ● extensive evaluation
- 23. Discussion ● Could static analysis be used to improve the approach? ● The paper focuses on computer vision problems. Would this approach work with other problems as well? ● Could this approach be altered to minimize energy consumption instead? ● How could the implementation be improved?