Rickard Brännvall presented the work with Joar Svartholm, Rickard Liikamaa and Jonas Gustavsson at the 6th Swedish Workshop on Data Science (SweDS18) in Umeå, Sweden on November 20 -21, 2018. http://sweds2018.cs.umu.se/

In a modern datacenter large amounts of electric energy is converted to heat by CPUs, memory and network equipment. Cooling units and fans remove heat from the server racks, which further adds to the consumption of electricity. It can be challenging to design automatic controls that use these resources efficiently to simultaneously minimize electricity costs and avoid damage to equipment caused by overheating. In a step towards building such a regulator a scaled down system of six servers with fans was used, analysing both model-based and model-free regulators using TensorFlow.

TensorFlow is popular for training deep neural networks in supervised-, unsupervised- or reinforcement learning. In this application it is used for calibrating a traditional time-series models and for controller design, taking advantage of its ability to express sequential models and pass them through its powerful automatic differentiator.

The particular model used is a so called grey-box model, inspired by the physical laws of heat exchange, but with all parameters obtained by optimization. First the model is encoded as a RNN and exposed to the time-series data via n-step PEM, leveraging TensorFlow’s ability to express rich multi-step cost functions.

In regulator design one’s preferences in terms of tolerance for deviation from set-point, smoothness of control signal and the economics of resource use are expressed in a precise cost function. A MPC type regulator was directly obtained from the model by writing such a cost function on the n-step predictions. The optimal control signal is then solved for approximately in real-time (on CPU). This allows online use for server-fan control. It was also compared with model-free regulators; first a PID controller with fixed constants calibrated against our model in TensorFlow. Finally,an autotuning regulator was explored that adjust its parameters on-line by gradient descent.