Reducing Data Center Energy Consumption via Coordinated Cooling and Load Management By: Luca Parolini, Bruno Sinopoli, Bruce H. Krogh from CMU Presentation: Liang Hao
Motivation REDUCING the ever growing electricity consumption in data centers COORDINATING cooling and load management which is now mostly independent
Previous Work Computational fluid dynamic models to optimize the delivery of cold air Optimal load-balancing policy Temperature-aware manner
Modeling
Modeling(1): Computational network Composed of servers nodes that interact through the exchange of workloads This layer interacts with the external world by exchanging jobs
Modeling(2): Thermal network
Modeling(3): Server nodes
Modeling(4): Server nodes
Modeling(5): CRAC nodes Tin, Tout, Tref If Tref <= Tin, Tout would tend to Tref Else Tout would tend to Tin pw = f(Tin, Tout)
Modeling(6): Environment nodes pw = 0 Tin, Tout
Modeling(7): Control Inputs Controllable variables: the computational workload exchange, the server node power states and the CRAC node reference temperature
CMDP Formulation In order to formulate our optimization problem as a finite CMDP we have to identify: a finite set X of states, a finite set A of actions from which the controller can choose at each step t = k *, a set P xay of transition probabilities representing the probability of moving from a state x to a state y when the action a is applied, and a function c : X £A ! R of immediate costs for each time step. The total cost over a given time horizon is the sum of the cost incurred at each time step.
CMDP Formulation
Server nodes n=3 CRAC node r=1 Environment node e=0 Discrete-time model with time step
CMDP Formulation: Simplification Server1 and server2 do not exchange tasks Ignore electricity consumption by server3, the scheduler The overall computational network workload exchange is reduced to the choice of the mean value of s
CMDP Formulation Quantize Tout and Tref
Solution Use the Markov Decision Process Toolbox for MATLAB to solve the CMDP problem
Simulation Results
What’s insight Build a model that can reflect the real problem and solve it using mature solutions To transform a real problem into a mathematical model, we quantize sequential variables into discrete ones