AI/ML-driven closed-loop automation using ONAP - Kate Hsuan
1. A ML driven closed
loop automation using
ONAP
KATE
2020/08/31
2. Outline
• Introduction
• Recap
• ONAP design time and runtime.
• Basic concept of closed-loop automation.
• Telemetry and data analysis.
• How the ML improve the efficiency of system reactions.
• A use case- intelligent power management (ONS2019).
• Conclusions.
3. Introduction
• ONAP- 一個專為Telco市場量身打造的community 和 opensource project.
• A Linux foundation project.
• Led by AT&T.
• ONAP 的重點
• The concept.
• Design time and runtime disaggregate.
• Automate everything to reduce human errors.
• Driver- Telemetry and Data analysis.
• Use case- An Intelligent power management closed loop automation.
• Engagement.
4. Design-time and Run-time
• 對service deployment 而言,design-time and run-time 分離帶來
的好處:
• For developer:
• 把程式寫好就好。
• 包好。
• 再來model 執行環境(刻粿模)。
• For operator:
• 拿到粿模,丟進機器。
• 放上材料開始印粿。
• Telco industry中員工的分工是很絕對的。
8. Concept of Closed-loop Automation
• 就像車神~壓到石頭會自己把車騎正
• 或扶正!咩撲QQ
• 讓系統可以感知到 Service 的狀態。
• 讓系統可以藉由系統的回潰自動修改Service 讓系統恢復正常。
Telemetry
Automation
system
ServicesFeedbacks
Alerts/ Events
Configurations
Logs and
improvements
Administrators
9. ONAP Closed-loop Automation Example (in reality)
VIM
App1 App2
DCAE Policy
SO
Multicloud
SDNC
APPC
Telemetry
Event
Actions
Configs
Real NFV commands
11. DCAE Architecture
DCAE platform
Consul
DCAE runtime
Cloudify
Policy
handler
Deployment
handler
Config
binding
service
PolicyCLAMP
DMaaP
SNMP/ VES
collector
HV-VES
TCA
(Threshold
analyzer)
xNFs
12. ML Brings the Benefits to Automation
• Traditional reactive prediction model.
• 頭痛醫頭,腳痛醫腳。
• 一直在對telemetry sampling 並評估下一段sampling period 的未來。
• 也許改改weight 可能很快。
• 但是也想想orchestration 所需要的時間。
• Proactive prediction and problem prevention.
• 預防頭痛,預防腳痛。
• 早早就把服務開好等著好戲上場(自動化)。
• 預先準備那些需要時間開啟的服務。
• 預先了解需求的成長決定投資的順序。
• 但我喜歡兩種和在一起使用。
13. ML Framework Integration to DCAE (Official)
Challenges
Training is an offline task. We don’t worry
about that.
Inference.
• 扯到很多東西。保證頭暈。
• Microservice management
• Route management
• Model management
Acumos
Inference micro services management
system.
SDC modeling 一切的東西。
14. ML Framework Integration to DCAE (土砲)
DCAE runtime
SNMP/ VES
collector
HV-VES
Inference
model
xNFs
DMaaP (Kafka/卡夫卡)
DB
conductor
DB
Offline
training
Model
publisher
Events
Policy
Events
Model
repo
想象中的admin UI
Telemetry data
15. Before Starting to Integrate.
• Behavior collection.
• 需要來點真正的資料,找到user behavior的pattern。
• 寫data training program,假裝自己是Data scientist.
• 對一對圖看看有沒錯誤。
• Measurement of system (xNF) performance under different
criteria.
• 建立baseline。
• 這樣才能正確計算出在每種configuration 系統應該要有的表現。
18. What We did. (for monitoring and analytics)
DMaaP (Kafka)
VES
collector
xNFs and bare metal
Streaming
engine
DB
conductor
DB
Offline
training
Model
repo
ONAP
Policy
Trigger actions and
control VIMs and xNFs.
Events
Model
Telemetry
data stream
Data logging
19. Compare to Traditional Reactive Method
Reactive
最簡單的方法就是用moving average 來
track。
斜率的應用。
For example
• 報告學姐,2266日線和月線黃金交叉~
(scale out 當韭菜)。
• 發生死亡交叉就關幾個VNF囉。
慢,VNF可能準備好了事件都過去了。
Proactive (Prediction)
Prevent from something.
提早開服務增加availability。
快,但有時會做白工。
要付出overhead
20. Conclusions
• ONAP- Telco grade opensource project.
• A design concept for service operation and automation.
• 使用它的架構可以思考一下自己的operation flow 到底要長什麼
樣。
• 圓的還是扁的~anyway.
• 捲起袖子加入吧!