EnergyEfficient Policies for RequestDriven Soft RealTime Systems

1
Energy-Efficient Policies for Request-Driven
Soft Real-Time Systems

• Cosmin Rusu, Ruibin Xu, Rami Melhem, Daniel Mosse
• Computer Science Department, University of
  Pittsburgh
• Euromicro Conference on Real-Time Systems
• (ECRTS) 2004

2
Outline

• Introduction
• System Model
• DVS policies
• Application-Oblivious
• Application-Aware
• Stochastic
• Experimental Results
• Conclusions

3
Introduction

• Power consumption is becoming the key design
  Issue
• DVS can fully explore power-performance tradeoffs
  
• For real-time systems with predictable workloads
• But, in most real-life situations, there is no
  a-priori knowledge of workloads such as arrival
  time and execution time of job/requests
• ? Workload is rather unpredictable!
• This work evaluates DVS algorithms for systems
  with unpredictable workloads
• Prediction-based schemes
• Stochastic schemes

4
System Model

• M discrete frequencies f1, f2, , fM
• Average power consumption is Pi at each freq. fi
• System power when idle is Pidle
• Arrival times and execution times for requests
• Not known beforehand
• However, deadline(soft deadline D) and type are
  known upon request arrival
• Scheduling of requests
• A first-come first-served(FCFS) fashion without
  preemption
• ? DVS algorithm is invoked when a timeout
  expires or upon
• arrival of a new request or completion of
  current request

5
DVS policies Application-Oblivious

• Interval-based algorithm
• System utilization is monitored every p time
  units, irrespective of application running
• If system was more than a utilization threshold
  during last period,
• Speed is increased to next higher discrete freq.
• If utilization u is less than a utilization
  threshold,
• Speed is set as
• Advantage simplicity
• Disadvantage fixed interval, unawareness of
  request deadlines

6
DVS policies Application-Aware

• Application-aware algorithm
• Predicts future performance needs by monitoring
  request inter-arrivals and execution times
  periodically
• Speed s is updated as
• Ni predicted no. of requests of type i
• ai predicted average execution time
• Li no. of requests of type i unfinished from
  previous period

7
DVS policies Stochastic

• Stochastic algorithm
• Data collected is probability distribution of
  request CPU cycles
• Building a histogram of cycles
• S granularity, WC worst-case no. of cycles of a
  request
• ,
• Cumulative density
• CDFk
• Our DVS scheme
• Chooses a primary speed fi and a secondary speed
  fj to minimize expected energy consumption
• If WCT/D gt fM, fi fj fM
• else if WCT/D ltf1, fi fj f1
• else compute energy for all combinations of fi
  lt WCT/D and
• fj gt WCT/D and select the pair with
  smallest energy

8
DVS policies Stochastic

• Stochastic algorithm(cont.)
• Computes ti, tj
• Expected energy consumption of bin k
• If (k1)S lt tifi,
• If (k1)S gt tifi,
• ? Total expected energy
• What is different from Ishiharas
• Considering probability distribution
• Primary and secondary speeds can differ by more
  than one discrete level
• If most requests are short enough to finish
  execution at primary speed, more savings will be
  resulted compared to adjacent speeds

9
Experimental Setup

• Power model
• Based on actual power measurements on IBMs
  PPC405LP
• Trace Data
• No.of cycles for each request in Event Extraction
  and CAF trace data from Mambo simulator
• Fixed bin width of 10 million cycles
• 76 bins for EE(753 million), 505 bins for
  CAF(5045 million)
• Parameter
• Period for monitoring system utilization, request
  inter-arrival time and processing time p1
  second
• Speed change overhead 1 millisecond

10
Experimental Results

• Stochastic approach achieves the most
  energy savings
• Up to 1/ 28.5 compared to no-power-management
• Up to 40 less energy compared to the second-best
  algorithm

11
Experimental Results(cont.)
Real trace(Event Extraction 81min.)
Synthetic trace
12
Conclusions

• Evaluated several DVS policies with unpredictable
  workloads
• Proposed a simple but effective stochastic DVS
  scheme
• This scheme achieved one order of magnitude
  energy reduction over no-power-management
• And up to 50 more savings over the best
  prediction-based scheme