We have developed a self-tuning feedback controller that caps the power consumed by a server to a user-defined budget. The controller’s parameters are tuned dynamically by computing an internal model of the server’s power consumption and how it changes with frequency. This allows the controller to adapt in response to changing workload and changing system configuration. The power model encompasses both clock modulation and voltage-frequency scaling actuators so they can be used together for power capping. We show the self-tuning controller achieves acceptable settling times and has application performance similar to a controller based on fixed parameters that are determined off-line. The selftuning controller saves the vendor time in determining good controller parameters and adjusts to unforeseen workloads and server configurations.
[1]
Zhikui Wang,et al.
Feedback Control Algorithms for Power Management of Servers
,
2008
.
[2]
Xiaorui Wang,et al.
Power capping: a prelude to power shifting
,
2008,
Cluster Computing.
[3]
Robert R. Wolford,et al.
BladeCenter chassis management
,
2005,
IBM J. Res. Dev..
[4]
Karthick Rajamani,et al.
A performance-conserving approach for reducing peak power consumption in server systems
,
2005,
ICS '05.
[5]
Xiaorui Wang,et al.
Cluster-level feedback power control for performance optimization
,
2008,
2008 IEEE 14th International Symposium on High Performance Computer Architecture.
[6]
Christoforos E. Kozyrakis,et al.
A Comparison of High-Level Full-System Power Models
,
2008,
HotPower.