SelfOptimization in Computer Systems via Online Control: Application to Power Management

1
Self-Optimization in Computer Systems via Online
Control Application to Power Management

• Presented by Dr. Ying Lu
• Feb. 6, 2006

2
Paper

• N. Kandasamy, S. Abdelwahed, and J. P. Hayes,
• Self-Optimization in Computer Systems via
  Online Control Application to Power Management
• Proceedings of International Conference on
  Autonomic Computing (ICAC04), New York, USA, May
  2004, pp. 5461.
• Self-Star Properties in Complex Information
  Systems, Lecture Notes  in Computer Science, vol.
  3460, Springer-Verlag, pp.174-189, 2005.

3
Contributions

• A generic control framework for self-optimizing
  computer systems
• Its application to minimize the power consumption
  of a single computer

4
Online Control Technique

• Starting from the initial system state
• Forecast the future system QoS behavior
• The sequence of control actions (a1, a2, a3,
  an) resulting in the best system behavior is
  chosen, a1 will be applied

x0
a1
x1j
n1
a2

n2
x2k

look-ahead prediction horizon
5
The Power Management Problem

• Assume a processor with a limited number of
  discrete frequency settings (such as AMD-K-2 or
  StrongARM)
• To satisfy QoS requirements of its workload while
  operating the processor at the lowest possible
  frequency
• How is this problem different from
• S Irani, S Shukla, R Gupta, Online Strategies
  For Dynamic Power Management In Systems With
  Multiple Power-Saving States, ACM Transactions on
  Embedded Computing Systems, 2003

6
Disk Power Management

• Dynamic Power Management (DPM)
• Transfer the Hard-Drive to lower power usage
  state when it is idle

7
CPU Power Management

• Dynamic Voltage Scaling (DVS)
• Multiple operating states CPU at different
  frequencies
• Transfer the CPU to the lowest frequency that can
  satisfy the request latency

8
Making the Control Decision

• What is the control decision?
• CPU frequency setting for the next time interval
• What are the factors for making the control
  decision?
• Latency Power Consumption
• Latency lt ?ref

9
Processor Models

• Latency
• c the time required to
• process request at fmax
• c/?(t), where ?(t)f(t)/fmax the processing time
  required at frequency f(t)
• Its reverse ?(t)/c the service rate ?(t)
• Power consumption
• E(t)?2(t)

10
The Online Controller Structure
11
Forecasting Workload

• Arrival Rate
• Processing Time

12
Whats Next
13
State Evaluation
14
(No Transcript)
15
Simulation

• Assume an AMD Athlon with 5 possible operating
  frequencies
• Workload arrival rate, requested objects, object
  popularity, object request frequency (temporal
  locality), processing time

16
Configurations

• Cost function w1100, w21
• For forecast models of
• arrival rate ?0.17, ?0.1
• request processing time ?0.35
• Do we need to adapt those parameters to deal with
  workload changes?

17
(No Transcript)
18
Effect of Prediction Horizon
19
Future Work

• Better (maybe adaptive) workload forecast
  mechanism?
• Those prediction mechanism from the Disk Power
  Management paper?
• Combining load-balancing, disk power management
  and CPU power management in a cluster computing
  environment?

20
Further Reading

• Y. Chen, A. Das, W. Qin, A. Sivasubramaniam, Q.
  Wang, and N. Gautam, Managing Server Energy And
  Operational Costs In Hosting Centers, In ACM
  SIGMETRICS, June 2005.

21
(No Transcript)