Journal
2017 FIFTH INTERNATIONAL SYMPOSIUM ON COMPUTING AND NETWORKING (CANDAR)
Volume -, Issue -, Pages 427-433Publisher
IEEE
DOI: 10.1109/CANDAR.2017.44
Keywords
Power Model; Workload Characterization; Power Estimation; Machine Learning
Ask authors/readers for more resources
In a power-aware scheduling system, power models are leveraged as the basis of estimating the effect of scheduling tasks. Previous studies showed that power consumption of servers is a non-linear function of resource usage, and a single set of parameters in one model can't accurately estimate power consumption at different usage levels. Gaussian Mixture Model (GMM) is a unsupervised machine learning model, which contains multiple GMM clusters. These clusters can be used to corelate power consumption with resource features at different usage levels. In this paper we further adapt GMM for power estimation in a distributed computing cluster. We use basic OS-reported resource features (CPU utilization, memory utilization etc.) of a server in our GMM, and this makes operators easily monitor and control the state of the server once scheduling decision is made. In addition, our GMM uses conditional probability to obtain fine-grained regression. We train the model using full features, which has the higher accuracy comparing with that only using CPU or part of features. In the end, we evaluate the power models in terms of not only accuracy but also usability. Compare to other linear or non-linear models, GMM has the highest accuracy but requires the longest training time.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available