Limits...
Power and Performance Management in Nonlinear Virtualized Computing Systems via Predictive Control.

Wen C, Mu Y - PLoS ONE (2015)

Bottom Line: Furthermore, because of improvement on chips and reduction of idle power, power consumption in modern machines shows significant nonlinearity, making linear power models(which is commonly adopted in previous work) no longer suitable.To deal with this, we build a discrete system state model, in which all control actions and time delay effect are included by state transition and performance and power can be defined on each state.Experiment results show the effectiveness of the controller.

View Article: PubMed Central - PubMed

Affiliation: Department of Computer Science and Engineering, Beihang University, Beijing, China.

ABSTRACT
The problem of power and performance management captures growing research interest in both academic and industrial field. Virtulization, as an advanced technology to conserve energy, has become basic architecture for most data centers. Accordingly, more sophisticated and finer control are desired in virtualized computing systems, where multiple types of control actions exist as well as time delay effect, which make it complicated to formulate and solve the problem. Furthermore, because of improvement on chips and reduction of idle power, power consumption in modern machines shows significant nonlinearity, making linear power models(which is commonly adopted in previous work) no longer suitable. To deal with this, we build a discrete system state model, in which all control actions and time delay effect are included by state transition and performance and power can be defined on each state. Then, we design the predictive controller, via which the quadratic cost function integrating performance and power can be dynamically optimized. Experiment results show the effectiveness of the controller. By choosing a moderate weight, a good balance can be achieved between performance and power: 99.76% requirements can be dealt with and power consumption can be saved by 33% comparing to the case with open loop controller.

No MeSH data available.


Linear model cannot reflect the reduced idle power of new machine.
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4520487&req=5

pone.0134017.g001: Linear model cannot reflect the reduced idle power of new machine.

Mentions: In most practical cases, linear power model is suitable and can provide convenience for utilization, see [8]. This causes nonlinear power models rare to be studied in the literature. However, linear power model has a defect, which can be seen from Fig 1. For the server Sun Netra x4250 and IBM x3450, power can be approximated by a linear function with respect to CPU load, where the positive power value when load equals zero denotes the idle power. However, when the machine is turned off, the system load is also zero, but the power value now is zero rather than idle power. Thus, linear power model cannot distinguish the state when the PM is shut down from the state when it holds zero workload, which makes turning on/off a PM deleted from the optional control action set in most previous work. As we have stated in Section 1, this does not fit the virtualized computing systems.


Power and Performance Management in Nonlinear Virtualized Computing Systems via Predictive Control.

Wen C, Mu Y - PLoS ONE (2015)

Linear model cannot reflect the reduced idle power of new machine.
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4520487&req=5

pone.0134017.g001: Linear model cannot reflect the reduced idle power of new machine.
Mentions: In most practical cases, linear power model is suitable and can provide convenience for utilization, see [8]. This causes nonlinear power models rare to be studied in the literature. However, linear power model has a defect, which can be seen from Fig 1. For the server Sun Netra x4250 and IBM x3450, power can be approximated by a linear function with respect to CPU load, where the positive power value when load equals zero denotes the idle power. However, when the machine is turned off, the system load is also zero, but the power value now is zero rather than idle power. Thus, linear power model cannot distinguish the state when the PM is shut down from the state when it holds zero workload, which makes turning on/off a PM deleted from the optional control action set in most previous work. As we have stated in Section 1, this does not fit the virtualized computing systems.

Bottom Line: Furthermore, because of improvement on chips and reduction of idle power, power consumption in modern machines shows significant nonlinearity, making linear power models(which is commonly adopted in previous work) no longer suitable.To deal with this, we build a discrete system state model, in which all control actions and time delay effect are included by state transition and performance and power can be defined on each state.Experiment results show the effectiveness of the controller.

View Article: PubMed Central - PubMed

Affiliation: Department of Computer Science and Engineering, Beihang University, Beijing, China.

ABSTRACT
The problem of power and performance management captures growing research interest in both academic and industrial field. Virtulization, as an advanced technology to conserve energy, has become basic architecture for most data centers. Accordingly, more sophisticated and finer control are desired in virtualized computing systems, where multiple types of control actions exist as well as time delay effect, which make it complicated to formulate and solve the problem. Furthermore, because of improvement on chips and reduction of idle power, power consumption in modern machines shows significant nonlinearity, making linear power models(which is commonly adopted in previous work) no longer suitable. To deal with this, we build a discrete system state model, in which all control actions and time delay effect are included by state transition and performance and power can be defined on each state. Then, we design the predictive controller, via which the quadratic cost function integrating performance and power can be dynamically optimized. Experiment results show the effectiveness of the controller. By choosing a moderate weight, a good balance can be achieved between performance and power: 99.76% requirements can be dealt with and power consumption can be saved by 33% comparing to the case with open loop controller.

No MeSH data available.