-
Notifications
You must be signed in to change notification settings - Fork 179
Roadmap
In the near term ( up to 12 months)
Kepler requires many moving parts: Kepler, Model Server, Model Estimator, and ServiceMonitor and Grafana dashboard for metrics collection and visualization.
All these can be easily handled by the Operator pattern. For that reason, the kepler-operator project is started to evaluate how to configure, install, and extend Kepler for other services by Kubernetes native patterns.
Kepler targets x86 platforms in the beginning. This is due to the prolific research on power measure on x86 platforms. However, as the we gain more visibility into workflow and methodology, we are aiming to support other platforms such as s390x and ARM64.
With more and more AI/ML workload, it is important to measure GPU power consumption. We are reviewing research papers and starting to improve the GPU power estimate models.
In the long term (12+ months)
Kepler is tested on bare metal environments and some Cloud platforms. Due to the diverse platforms on different Cloud platforms, more testing and qualification are needed to support broad hybrid Cloud environments.
The key goal of running ML models for power estimate is to improve the accuracy. The modeling process and power models are being constantly tested, validated, and improved towards that goal.
Kepler can help many other use cases in the CNCF ecosystem. For instance, projects such as PEAKS and CLEVER use Kepler metrics for energy efficiency aware Kubernetes Pods scheduling and scaling, respectively.