Scaling and Shifting CPU and GPU use to Reduce Energy Consumption and Peaks for Data Centers

Purpose and goal:

The goal of this project was to reduce the energy consumption and power peaks of applications running on data centers, by dynamic scaling of resource allocations to actual needs. Users can also shift usage from electricity load peak hours to times of less utilization, such as nighttime. The project has fulfilled its goals, and Elastisys together with Elastx can now offer auto-scaling with both CPUs and GPUs, with the potential of significant energy savings.

Expected results and effects:

To the best of our knowledge, the joint commercial offer of Elastisys and ElastX, that was developed as a direct result of this project, is the first one available in Sweden that combined Kubernetes as a Service with GPUs and auto-scaling. This means that Swedish customers can now run machine learning workflows with data staying in Swedish jurisdiction: cheaply, energy efficiently and with good developer experience.

Approach and implementation:

The project demonstrates the benefits of open source and cloud native technologies. These building blocks have recently matured significantly, and today allows relatively small Swedish cloud providers to create best of breed commercial offers. Specifically, we were positively impressed with improvements in the Kubernetes Cluster API provider for OpenStack as well as the Nvidia GPU operator. The collaboration was smooth as both Elastisys and ElastX are committed to open source and cloud native.