![State-of-the-art system
StreamApprox [Middleware’17]
Approximate output
± error bound
StreamApprox
Stream
aggregator
S1
S2
Sn
…
Data
stream
Cloud datacenter
Limitations:
• It wastes bandwidth
• It utilizes only cloud datacenter resources](https://image.slidesharecdn.com/edgecomp-221121230705-0271fb4c/85/Edge-Comp-pptx-6-320.jpg)
Edge Comp.pptx
- 1. ApproxIoT Approximate Analytics for Edge Computing https://ApproxIoT.github.io/ApproxIoT/ Zhenyu Wen, Do Le Quoc, Pramod Bhatotia, Ruichuan Chen, Myungjin Lee
- 2. Modern online services Stream aggregator Stream analytics system Useful Information Processing streaming data from different sources
- 3. Modern online services Low latency Tension Approximate computing Efficient resource utilization
- 4. Approximate computing Many applications: Approximate output is good enough! The proportion of data is useful for this application Live taxi heatmap
- 5. Approximate computing Idea: To achieve low latency, compute over a sub-set of data items instead of the entire data-set Analyze Approximate output ± error bound Approximate computing (sampling)
- 6. State-of-the-art system StreamApprox [Middleware’17] Approximate output ± error bound StreamApprox Stream aggregator S1 S2 Sn … Data stream Cloud datacenter Limitations: • It wastes bandwidth • It utilizes only cloud datacenter resources
- 7. Edge computing Cloud Gateway Edge node Local processing Source of data Allows data to be processed at the edge node before it’s sent to the cloud Opportunities: • Providing more computing resources • Saving bandwidth
- 8. Edge infrastructure Source: https://peering.google.com/#/infrastructure Azure IoT edge Watson IoT AWS IoT
- 9. Problem statement To build a stream analytics system • By utilizing the cloud and edge computing resources • By leveraging approximate computing Design goals • Efficiency: Efficient utilization of computing resources • Adaptability: Adaptive execution based on the available resources • Transparency: No code change required and resource management
- 10. Outline • Motivation • Design • Implementation • Evaluation
- 11. ApproxIoT: Overview S1 Si Sn … Sm … … Central node Cloud Query Approximate output ± error bound ApproxIoT ApproxIoT employs sampling in the distributed environment of edge + cloud Edge nodes Regional edge Continental node
- 12. Naïve algorithm SRS Query Simple random sampling (SRS) Approximate output ± error bound Sampled unfairly Overlooked Low accuracy
- 13. Background: Stratified sampling Stratified sampling Advantage: The sub-streams are sampled fairly Disadvantage: Requires the knowledge of each sub-stream size
- 14. Background: Reservoir sampling Reservoir sampling Size of reservoir = 4 Reservoir sampling Size of reservoir = 4 Advantage: • No pre-knowledge required of sub-stream size Disadvantages: • The sub-streams are sampled unfairly • Difficult to run on multiple nodes Reservoir sampling Size of reservoir = 4 The 5th item With probability( 4 5 ) replaced by the 5th item Reservoir sampling Size of reservoir = 4 Reservoir sampling Size of reservoir = 4 The 6th item With probability( 4 6 ) replaced by the 6th item Reservoir sampling Size of reservoir = 4 Reservoir sampling Size of reservoir = 4
- 15. ApproxIoT sampling algorithm Easy to parallelize, requires no synchronization between sub-streams Weighted hierarchical sampling (WHS) Combining stratified and reservoir sampling Weight: C/N, if C>N 1, if C <=N WHS Reservoir size N=4 With initial weight 1 W=1 W=1 W=1 W=6/4 W=1 W=1 C=6
- 16. WHS on edge nodes Regional edge WHS W=1 W=1 W=1 W=6/2=3 W=4/2=2 W=1 Continental node WHS W=4 W=1 W=3 W=4*5/2=10 W=1*3/2=3/2 W=3 Reservoir size equals 2 Central node Cloud Edge nodes Regional edge Continental node Easy to parallelize, requires no synchronization between computing nodes Carried weight Current weight
- 17. ApproxIoT in the cloud Reservoir size equals 1 Query (sum) WHS The weights are carried W=4/3*6/1 =8 W=1*4/1=4 W=1*2/1=2 ± error bound 8* +4* +2* W=4/3 W=1 W=1 Approximate output: Central node Cloud Edge nodes Regional edge Continental node
- 18. Outline • Motivation • Design • Implementation • Evaluation
- 19. Implementation S1 S2 Sn … Kafka cluster Stream pub/sub Edge nodes Cloud datacenter Data stream Sampled data stream Sampled data stream See the paper for more details Kafka Streams
- 20. Experimental setup • Evaluation questions • Accuracy vs. sample size • Throughput vs. sample size • Testbed: 25 nodes • 15 nodes for ApproxIoT deployment • 10 nodes for Kafka cluster • Datasets: • Synthetic: Poisson and Gaussian distribution • Real: Brasvo pollution and New York Taxi Ride See the paper for more results!
- 21. Accuracy vs. sample size 0 20 40 60 80 10 20 40 60 80 Accuracy loss(%) Sampling fraction(%) SRS ApproxIoT Lower the better ApproxIoT: ~2600X higher accuracy over SRS The average is 0.035%
- 22. Throughput vs. sample size 0 40 80 120 10 20 40 60 80 90 100 Throughput(k) items/s Sampling fraction(%) Native SRS ApproxIoT Higher the better • ApproxIoT has low overhead compared to the native execution • ApproxIoT has similar throughput as SRS
- 23. Conclusion ApproxIoT: Approximate analytics for edge computing Adaptability Adaptive execution based on the available resources Transparency Requires no code changes and resource management Thank you! More details on the project website: https://ApproxIoT.github.io/ApproxIoT/ Efficiency Efficient computing and bandwidth resource utilization