An Edge Datastore Architecture For Latency-Critical ... · Arun Ravindran and Anjus George UNC...
Transcript of An Edge Datastore Architecture For Latency-Critical ... · Arun Ravindran and Anjus George UNC...
![Page 1: An Edge Datastore Architecture For Latency-Critical ... · Arun Ravindran and Anjus George UNC Charlotte 1. Distributed Vision at the Edge - Smart City Warn pedestrian about potential](https://reader034.fdocuments.us/reader034/viewer/2022052022/6036eb316667ce7ac8345659/html5/thumbnails/1.jpg)
An Edge Datastore Architecture For Latency-Critical Distributed Machine
Vision ApplicationsArun Ravindran and Anjus George
UNC Charlotte
1
![Page 2: An Edge Datastore Architecture For Latency-Critical ... · Arun Ravindran and Anjus George UNC Charlotte 1. Distributed Vision at the Edge - Smart City Warn pedestrian about potential](https://reader034.fdocuments.us/reader034/viewer/2022052022/6036eb316667ce7ac8345659/html5/thumbnails/2.jpg)
Distributed Vision at the Edge - Smart City Warn pedestrian about potential accidents
Automatically detect and alert drunk driving
Effective bias free law enforcement
Source: YouTube 2
![Page 3: An Edge Datastore Architecture For Latency-Critical ... · Arun Ravindran and Anjus George UNC Charlotte 1. Distributed Vision at the Edge - Smart City Warn pedestrian about potential](https://reader034.fdocuments.us/reader034/viewer/2022052022/6036eb316667ce7ac8345659/html5/thumbnails/3.jpg)
System Architecture
Cameras
End nodes
Edge Servers
Cloud
3
![Page 4: An Edge Datastore Architecture For Latency-Critical ... · Arun Ravindran and Anjus George UNC Charlotte 1. Distributed Vision at the Edge - Smart City Warn pedestrian about potential](https://reader034.fdocuments.us/reader034/viewer/2022052022/6036eb316667ce7ac8345659/html5/thumbnails/4.jpg)
Vision Edge Datastore ● Applications at Edge
analyzes data collected by End nodes to detect events○ Need data store that persists data
gathered from multiple end nodes○ Able to specify latency required
● Challenge - how to maintain low latency at edge ?○ Latency sources - wireless
channel, bufferbloat, read/write latency 4
Latency CDF witn node scaling
![Page 5: An Edge Datastore Architecture For Latency-Critical ... · Arun Ravindran and Anjus George UNC Charlotte 1. Distributed Vision at the Edge - Smart City Warn pedestrian about potential](https://reader034.fdocuments.us/reader034/viewer/2022052022/6036eb316667ce7ac8345659/html5/thumbnails/5.jpg)
Cloud vs Edge● Data Center vs. “Field”
○ Security, Fault tolerance
● Wired vs. Wireless○ Bandwidth, latency
● Homogeneous vs. Heterogeneous○ ARM/x86 SoCs, Multiple storage and networking technologies
● Distributed data storage vs. Distributed data at source○ Big, fast, distributed data○ latency critical/sensitive applications
5
![Page 6: An Edge Datastore Architecture For Latency-Critical ... · Arun Ravindran and Anjus George UNC Charlotte 1. Distributed Vision at the Edge - Smart City Warn pedestrian about potential](https://reader034.fdocuments.us/reader034/viewer/2022052022/6036eb316667ce7ac8345659/html5/thumbnails/6.jpg)
Prior Work at Edge Storage
● VisFlow Project (Microsoft)● PathStore Project (Toronto)● Cachier Project (CMU)
6
![Page 7: An Edge Datastore Architecture For Latency-Critical ... · Arun Ravindran and Anjus George UNC Charlotte 1. Distributed Vision at the Edge - Smart City Warn pedestrian about potential](https://reader034.fdocuments.us/reader034/viewer/2022052022/6036eb316667ce7ac8345659/html5/thumbnails/7.jpg)
Our Design philosophy at Edge
● Application specific systems○ Tension between specificity and generality
● Autonomous operation○ Techniques from Control Theory and AI (ML, Deep Learning,
Reinforcement Learning)
7
![Page 8: An Edge Datastore Architecture For Latency-Critical ... · Arun Ravindran and Anjus George UNC Charlotte 1. Distributed Vision at the Edge - Smart City Warn pedestrian about potential](https://reader034.fdocuments.us/reader034/viewer/2022052022/6036eb316667ce7ac8345659/html5/thumbnails/8.jpg)
Key idea - Exploit application characteristics
● Two type of data - image feature vectors (1-10 kB) and image keyframes (100 - 500 MB)
● Feature vectors - latency critical○ Tracking, behavioral analysis
● Keyframes - latency sensitive○ Archival
● Feature vector latency by sacrificing keyframe accuracy○ Need to do this dynamically since channel interference and scene
content is dynamic
8
![Page 9: An Edge Datastore Architecture For Latency-Critical ... · Arun Ravindran and Anjus George UNC Charlotte 1. Distributed Vision at the Edge - Smart City Warn pedestrian about potential](https://reader034.fdocuments.us/reader034/viewer/2022052022/6036eb316667ce7ac8345659/html5/thumbnails/9.jpg)
Key idea - Latency control knobs
● Control knob 1: Keyframe TX ○ Controls the rate at which keyframes are transmitted○ Low egress rates could result in bufferbloat
● Control knob 2: Keyframe Sim○ Drops similar keyframes to maintain buffer length○ Accuracy vs. Latency trade off○ Needs a similarity metric
9
![Page 10: An Edge Datastore Architecture For Latency-Critical ... · Arun Ravindran and Anjus George UNC Charlotte 1. Distributed Vision at the Edge - Smart City Warn pedestrian about potential](https://reader034.fdocuments.us/reader034/viewer/2022052022/6036eb316667ce7ac8345659/html5/thumbnails/10.jpg)
Vision Edge Data Store - Design
● End node processing generates key frames feature and feature vectors
● Inserted with timestamp and node ID into transmit buffer● Data transmitted to Edge server● Aggregate and persist data at Edge server
○ Low latency store (RocksDB, RAMCloud)
● End node controls keyframe Tx rate and buffer length○ Scalable since controllers are independent
10
![Page 11: An Edge Datastore Architecture For Latency-Critical ... · Arun Ravindran and Anjus George UNC Charlotte 1. Distributed Vision at the Edge - Smart City Warn pedestrian about potential](https://reader034.fdocuments.us/reader034/viewer/2022052022/6036eb316667ce7ac8345659/html5/thumbnails/11.jpg)
Prototype Evaluation Platform● Emulation platform
○ LXC containers for nodes○ NS3 network simulator for WiFi channel○ Client/Servers implemented in Golang○ Image similarity (SSIM) with Python sckit-image○ qperf for latency measurements
● Controller○ Bang-bang (on/off) control
● Data○ 500kB keyframes, 4 kB feature vectors○ External interference simulated via Poisson process (5s TX, λ = 30s)
11
![Page 12: An Edge Datastore Architecture For Latency-Critical ... · Arun Ravindran and Anjus George UNC Charlotte 1. Distributed Vision at the Edge - Smart City Warn pedestrian about potential](https://reader034.fdocuments.us/reader034/viewer/2022052022/6036eb316667ce7ac8345659/html5/thumbnails/12.jpg)
Results
12
Latency CDF - Keyframe TX control Latency CDF - Keyframe Sim control
![Page 13: An Edge Datastore Architecture For Latency-Critical ... · Arun Ravindran and Anjus George UNC Charlotte 1. Distributed Vision at the Edge - Smart City Warn pedestrian about potential](https://reader034.fdocuments.us/reader034/viewer/2022052022/6036eb316667ce7ac8345659/html5/thumbnails/13.jpg)
Keyframe similarity (SSIM) - Pedestrian crash video
Accuracy vs. Latency tradeoffs
13
![Page 14: An Edge Datastore Architecture For Latency-Critical ... · Arun Ravindran and Anjus George UNC Charlotte 1. Distributed Vision at the Edge - Smart City Warn pedestrian about potential](https://reader034.fdocuments.us/reader034/viewer/2022052022/6036eb316667ce7ac8345659/html5/thumbnails/14.jpg)
On going work
● Experimental characterization of interference, keyframe similarity, application requirements
● Internal interference - scheduling problem?○ Distributed - client-server vs. peer-to-peer○ Dependence on scene dynamics
● Control / Learning algorithms
14
![Page 15: An Edge Datastore Architecture For Latency-Critical ... · Arun Ravindran and Anjus George UNC Charlotte 1. Distributed Vision at the Edge - Smart City Warn pedestrian about potential](https://reader034.fdocuments.us/reader034/viewer/2022052022/6036eb316667ce7ac8345659/html5/thumbnails/15.jpg)
Request Feedback
● On use of WiFi at Edge for latency critical apps? ● On differences between cloud and edge storage?● What would you like to see experimentally validated?● How should latency/accuracy requirements be
communicated from Edge app. to camera end nodes?● Are there other edge applications that are similar?● What edge specific security issues should we consider?● Any experience with simulating NS3 802.11ac with
containers? 15