autonomous datacenter module machin learning

The ITEA3 AutoDC project is ending with a final review

10 October 2021, 14:48

The AutoDC project has been running for 3 years. It has been funded via the EUREKA cluster ITEA3 by national funding agencies eg. Vinnova in Sweden.

The aim of AutoDC was to provide an innovative design framework for autonomous data centers. An autonomous datacenter should be able to, without any human intervention, from a best effort perspective continue its operation independent of contextual interference, such as intermittent power failure, failing components, overheating etc.

The vision of the project was “Bring the datacenter container to site, leave it for 5 years and then exchange the complete datacenter”. Since the project had limited time 3 years and budget it could not test a full vision but the goal was to develop technology enabler to be able to achieve the vision.

The partners in the project came from three countries. From Canada Mariner, MLT and Ericsson Canada with St. Marys university. From Finland Granlund, kW-Set, Orbis and Aalto university. From Sweden the project coordinator Ericsson, Swegon, Comsys, ITRS, Hi5 (Advania) Swedish Modules, Clavister, KTH, LTU, Lund university and RISE ICE Datacenter.

During the final review the autonomous datacenter was presented from different angels. MLT and Granlund described requirements on the datacenter envelope and software and also compared ROI for different designs. kW-Set, Swegon, Comsys and RISE demonstrated improvements to infrastructure components back-up power, cooling and UPS for longer life-time and less maintenance. Less moving parts will be important. Both ITRS and Clavister showed data collection, monitoring, security and analysis improvements in their products on the way autonomous operation. Hi5 showed their automatic resource-invoicing system based on data from Comsys and ITRS.

An important part of an autonomous operation is the machine learning and automatic control that is needed. Ericsson described their work on transfer learning used to update ML models and their RL methods for optimization together with Lund university, Mariner described the complete software architecture needed with all parts from data collection to control, RISE showed risk aware control of a edge node together with LTU, Aalto showed work on state discovery and prediction and KTH described their work on online feature selection.

All in all, the project successfully developed technology enablers and showed during the final review steps taken and methods needed to reach zero-touch operations. This was a first major step towards the vision but more work is needed to reach the envisioned objectives.

https://autodc.tech

Jonas Gustafsson

Research program manager

+46 70 333 00 35

Tor Björn Minde

Enhetschef

+46 70 624 29 59