A term paper on

1

• A term paper on
• System Level Power Estimation
• And Optimization
• References
• Luca Benini,Robin Hodgson and Polly Siegel,
  System-level Power Estimation And Optimization,
  ISLPED 98, August 10-12, 1998, Monterey, CA USA.
• Luca Benini,Alessandro Bogliolo,and Giovanni De
  Micheli,A Survey of Design Techniques for
  System-Level Dynamic Power Management.
• Presented By
• Mukesh Agarwal 2003JVL0006

2
Introduction

• Power reduction is taking on increasing
  importance .
• As clock speeds increase, power dissipation and
  accompanying thermal heat dissipation problems
  increase.
• Portable systems require long battery life and
  low weight.
• Hence, power efficient designs at system level
  become important.
• A simple but powerful model is required for
  describing power behavior at system level.
• Efficient power management approaches to be
  implemented at system level.

3
Requirements of Power Model at System Level

• Fast estimation of system-level power useful for
  component selection and system partitioning
  phase.
• Supports high-level abstract system descriptions.
• Model should include
• Power behavior of system and its functional
  blocks.
• Block interactions and information about the
  environment.
• Abstracts away everything except power behaviors.

4
System Level Power Model

• Consists of
• A set of resources represented as power state
  machines.
• An environmental workload specification.
• A power manager that implements power
  management and control policies.

5
Basic Resource PSM

• Resources power behavior captured by a Power
  state machine.
• Different power states each with a
• fixed power consumption.
• Transitions among power states
• determined by power manager
• Include Performance costs and transition
  penalties during
• switching between power states.

6
PSM With Transition Penalties

• Transition penalties associated with changes from
  one power state into the next. e.g.
• - CPU component from the SLEEP state to IDLE
  state
• - Restarting disk drive involves time and power
  penalties
• Power and service annotation on states
• Power and delay annotation on edges

7
Modeling Utilization-dependent Resources

• The power consumption of some resources(e.g.
  memories) activity level dependent.
• Activity levels defined for each state.
• Activity level is specified by power manager .
• Can be changed while in same state.
• Psactivity_level peak_power.

8
Power Manager

• Implements power management policy.
• Decides which resource to send a command to
• and when to issue a command to a component.
• Issues power state change requests in response
• to requests from the external environment
• and internal resources.
• Policies to consider transition penalties and
  trade-offs between power and performance by
  adapting to actual workloads Dynamic power
  management
• Power consumption of power manager is negligible

9
Power Estimation

• Power model is mapped into an event-driven
  simulation paradigm .
• Model implemented using behavioral simulation
  language such as VHDL.
• Event-driven behavioral simulator used to run the
  model .
• Total power dissipated at any time derived by
  summation of power consumption of components in
  respective states and activity levels

10
Dynamic Power Management

• Achieve energy-efficient computation by
  selectively turning off or reducing the
  performance of components when idle or partially
  unexploited.
• Based on assumptions that
• Systems (and their components) experience non
  uniform workloads
• It is possible to predict the fluctuations in
  workload.
• Uses a control procedure/policy based on
  observations and/or assumptions on the workload.
• There exist different approaches to system-level
  DPM.
• Predictive schemes
• Stochastic optimum control.

11
The Applicability of DPM

• State transition power (Ptr) and delay (Ttr)
  Ptr Ttr
• If Ttr 0, Ptr 0 the policy is trivial
• - Stop a component when it is not needed
• Ptr Ttr
• If Ttr ! 0 or Ptr ! 0 (always)
• Shutdown only when idleness is long enough to
  amortize the cost
• What if T and P are not deterministic?

ON
OFF
12
System Break-even-time

• Minimum idle time for amortizing the cost of
  component shutdown
• TBE Ttr Ttr( Ptr Pon) / (Pon Poff)

Transition delay (Ttr)
Transition power (Ptr) Sleep
power (Poff)
13
When to Use Power Management

• When TBE lt T idle
• -Average idle periods are long enough
• -Transition delay is short enough
• -Transition power is low enough
• -Sleep power is low enough
• When designing system for a known workload

14
Predictive Schemes

• DPM decisions taken based on uncertain
  predictions.
• Exploits the correlation between the past history
  of the workload and its near future.
• Aim at predicting idle periods long enough to go
  to sleep modes ie, pTidle gt TBE .
• Go to sleep state if Tpred is long enough to
  amortize state transition cost.
• Main Issueprediction accuracyOver/under
  predictions.

15
Fixed Timeouts

• Follows simple policy
• If Tidle gt TTO go to SLEEP
• Stay in sleep until workload ! 0
• Rationale
• When Tidle gt TTO it is likely that T idle gt
  TTO TBE
• Choice of TTO is critical
• Large is safe, but it could be useless
• Too small is highly undesirable
• Limitations
• Performance penalty for wake-up is paid after
  every shutdown
• Power is wasted during TTO
• Apply predictive shutdown policies.

16
Predictive Shutdown

• Two approaches suggested
• A nonlinear regression equation obtained from
  the past history
• and used to make predictions.
• If Tpred gtTBE , the system is shut down as soon
  as it becomes idle.
• Eliminates power waste caused by TTO.
• 2. Duration of the busy period preceding the
  current idle period is observed.
• If o the idle period is
  assumed to be larger than TBE and the system is
  shut down.
• Short active periods are often followed by long
  idle periods.

17
When to Use Predictive Techniques

• When workload has memory
• Implementing predictive schemes
• Predictor families must be chosen based on
  workload types
• Predictor parameters must be tuned to the
  instance-specific workload statistics
• When workload is non-stationary or unknown,
  on-line adaptation is required.

18
Stochastic Control Approach

• Recognize inherent uncertainty
• Exact prediction of future events is
  impossible
• Abstraction of system model implies
  uncertainty
• Model components,system and workload as
  stochastic processes
• Expected values of cost metrics are optimized

19
Controlled Markov processes

• System and environment modeled as Markov chains
• -System is called service provider (SP)
• -Environment is called service requester (SR)
• SP is a controlled Markov chain
• -State transition probabilities depends on
  commands
• Cost metrics associate power and performance
  values with each system state-command
• The power manager (PM) observes the state of the
  system and issues commands to control evolution

20
Implementations of DPM

• Shutdown idle components
• Gate clock of idle units
• Clock setting and voltage setting
• Support multiple-voltage multiple-frequency
  components
• Components with multiple working power states
• Operating system-based power management
• The OS knows of tasks running and waiting
• The OS should perform the DPM decisions

21

Conclusions

• System-level energy-efficient design is an area
  of research with many opportunities
• A model based on behavioral language and event
  driven simulation engine derived.
• DPM is a powerful methodology for power efficient
  designs
• State of operation of various components is
  adapted to the required performance level.
• Problem of designing power management policies
  that minimize power under performance constraints
  is a challenging one.