PTSM- #1 - Monitoring, Data Center et Machine Learning

Support de présentation de Clément Bataille et Christophe Rannou sur l'usage du machine learning appliqué aux séries temporelles chez OVHCloud dans le cadre du monitoring de leur datacenters.

Transcript

1. Monitoring, Data Center et Machine Learning https://labs.ovh.com/machine-learning-platform Clement Bataille Christophe Rannou
2. WE OPERATE IN A FAST GROWING MARKET where availability is essential PUBLIC CLOUD & VPS PRIVATE CLOUD DEDICATED SERVER STORAGE NETWORK & SECURITY MULTI-LOCAL EVERYWHERE RAPID DELIVERY WORD-CLASS STANDARDS
3. OUR STORY
4. WE INNOVATE WITH EVERYONE WE ARE INDEPENDANT WE ARE THE ALTERNATIVE WHICH FREES UP THE CLOUD THROUGH CREATIVITY AND AN ENTREPRENEURIAL SPIRIT EMPLOYEES CUSTOMERS ECOSYSTEM OVH PARTNERS INNOVATION FOR FREEDOM
5. WE ARE A GLOBAL CLOUD PROVIDER WE ARE AN INDUSTRIAL PLAYER Hillsboro x1 17 Tbps 34 Point of Presence IN 3 CONTINENTS BHS x6 Vinthill x1 +1 400 000 CUSTOMERS IN 132 COUNTRIES Singapore x1 2 200 EMPLOYEES IN 18 COUNTRIES Sydney x1 Europe x18 BUILDER OF ITS OWN SERVERS SINCE 2002 + 1 million SERVERS built since 1999 29 DATA CENTERS in 12 locations EXISTING SINCE 1999
6. SIMPLE AND QUICK TO SET UP MULTI-LOCAL, CLOSE TO EACH PERSON THROUGHOUT THE WORLD WHOSE COST IS ACCESSIBLE AND PREDICTABLE REVERSIBLE, OPEN AND INTEROPERABLE TRANSPARENT AND RESPONSIBLE WE BUILD A DIFFERENT CLOUD a truly Cloud SMART TO GROW AND SUCCEED TOGETHER
7. ML for Time Series with Prescience 7
8. Toy Use Case 8 We are going to have a look at the international airline passengers prediction problem. The initial data is in the csv format : steps passengers 1949-01 112 1949-02 118 1949-03 132 1949-04 129
9. Problem Definition Detection Time Series • Specific Target • Ground Truth quality • Specific Target • Ground Truth quality • Forecast • Anomaly detection • Prediction horizon • Forecast • Anomaly detection • Prediction horizon 9
10. Problem definition 10 What kind of problem do you want to solve ? • Simple Classification • Simple Regression • Time series classification • Time series regression/forecast
11. Problem definition 11 • What is your label ? • What is your ordering column ? Month Passengers 1949-01 112 1949-02 118 1949-03 132 1949-04 129 Ordering column Label
12. Problem definition 12 • How many steps forward do you want to predict ? One year => 12 steps • Do you want to apply a discount ? No => 1 • Scoring metric ? Default => MSE 1 Year
13. Experiment definition 13 How many fold do you want to create ? 2
14. Experiment definition 14 How much budget to allow on optimization ? 10
15. Demo 15
16. Scores are deceitful • Data Leakage – Unusable – Dangerous – Difficult to detect – Difficult to correct • Wrong problem definition 16
17. Stock Value Prediction: How to get rich in 5 minutes Graphs: Vegard Flovik, How (not) to use Machine Learning for time series forecasting: Avoiding the pitfalls 17
18. Stock Value Prediction: Why I am not rich 18
19. Stock Value Prediction: What I should have done 19
20. Use cases
21. Electricity Comsuption forecast Sonde DC Site EDF WARP10 ML with Prescience Retrain model every week Forecast everyday
22. CEPH Key indicator: Forecast every 6h
23. VOIP Monitoring Chiffres Clés : 125 destinations 3 lignes par pays 1 point à l’heure
24. Temperature Anomaly Room Warp10 Arima Model Retrain every week Forecast by point Compare with Isolation Forest Alerting with TAT
25. Temperature Anomaly Rack Warp10 Mobil Mean Isolation Forest Alerting with TAT
26. Network forecast Key indicator : • 20K equipment • 100 metric by equipment • Point every minute
27. Network forecast • Trend Stationary • Differency Stationary Method stationary transformation Graphs: Investopedia, stationary and non-stationary processes : https://www.investopedia.com/articles/trading/07/stationary.asp
28. Network forecast Warp10 Heuristique de selection Regression Arima by TS Export to WARP10 + Mailing
29. T H A N K YO U @ChrisRannou Clement Bataille https://gitter.im/ovh/ai