Conservative Scheduling : Using Predicted Variance to Improve Scheduling Decisions in Dynamic Enviro

1
Conservative Scheduling Using Predicted
Variance to Improve Scheduling Decisions in
Dynamic Environments

• Lingyun Yang, Jennifer M. Schopf, Ian Foster
• University of Chicago
• Argonne National Laboratory

2
What is the problem?

• In multi-user time-shared systems
• Heterogeneity performance of resources are
  irregular
• Dynamicityresource sharing causes dynamicity
• Loosely synchronous iterative applications
• Tasks communicate between iterations
• Next iteration on a given resource cant begin
  until the communication to that resource has been
  finished

Variation causes delay
3
To use resources efficiently

• Require new approaches capable of dealing with
  heterogeneous and dynamic nature of such systems
• Our solution
• Time balancing scheduling strategy --
  heterogeneous
• Prediction of mean and variance CPU capability
  --dynamic
• Conservative data allocation strategy
  

4
Time balancing

• Assume the target set of resources is fixed
• Time balancing load are balanced between
  processors so that each processor finishes
  executing roughly at the same time
•    Di data assigned to processor i
•    Dtotal the total amount of data
•   Ei(Di) the execution time of task on
  processor,
• parameterized by
• Amount of data allocated
• Resource capabilityinfluenced by effective CPU
  load

5
To Decide Effective CPU Load

• What kind of future information do we need?
• CPU load at some future time point
• Average CPU load for next time interval
• Variation of CPU load over next time interval
• How to get these measures of future capability?
• One step ahead predictor
• Interval load prediction
• Load variation prediction
• How to use these future information to guide data
  mapping?
• Translating this measure into an effective
  resource capability that is then used to guide
  data mapping

6
Outline

• Problems
• gtWhat information do we need?
• CPU load at some future time point
• Average CPU load for next time interval
• Variation of CPU load over next time interval
• Three Predictors
• Conservative Scheduling Method
• Experiments
• Conclusion

7
What future information do we need

• CPU load at some future time point
• Not enough for most applications
• Average load for the next time interval
• Really needed for better data distribution and
  scheduling
• Variation of the load for the next time interval
• Avoid the wave of delayed behavior caused by
  variation

8
Outline

• Problems
• What information do we need?
• gtThree Predictors
• One-step-ahead CPU load prediction
• Interval load prediction
• Load variation prediction
• Conservative Scheduling Method
• Experiments
• Conclusion

9
One-Step-Ahead CPU load prediction

• Tendency-based time series predictor
• Cc1,c2cn the preceding CPU load time series
  measured at constant-width time interval
• pn1 the predicted value for measurement value
  cn1.

10
Interval load prediction

• Calculate an interval CPU load time series
• ai average CPU load over the time interval that
  is approximately equal to the application
  execution time
• Interval load prediction

11
Load variation prediction

• Calculate the standard deviation time series
• Si the average difference between the CPU load
  and the mean Load over the interval
• Load variation prediction

12
Outline

• Problems
• What information do we need?
• Three Predictors
• gtConservative Scheduling Method
• Experiments
• Conclusion

13
Conservative scheduling method

• Assign less work to less reliable
  (higher-variance) resources.
• Avoid the wave of delayed behavior caused by
  variance in the resource capability
• Conservative load prediction
• Effective CPU loadpredicted Meanpredicted
  Variance
• pa n1 ps n1

14
Other scheduling options

• One-Step Scheduling (OSS)
• Effective CPU loadpc n1
• Predicted Mean Interval Scheduling (PMIS)
• Effective CPU load pa n1
• History Mean Scheduling (HMS)
• Effective CPU load mean of the
  history load for the 5 minutes preceding the
  application start time
• History Conservative Scheduling (HCS)
• Effective CPU loadmean variance
  of the history load collected for 5 minutes
  preceding the application run

15
Outline

• Problems
• What future information do we need?
• Three Predictors
• Conservative Scheduling Method
• gtExperiments
• Experimental Methodology
• Results
• Result evaluation
• Conclusion

16
Experimental methodology

• ApplicationCactus
• Compare different methods fairly load trace
  Playback tool generates a background workload
  from a trace of the CPU load.
• 64 real load traces with different mean and s.d.
• Different application execution time
  
• 1 minutes 10 minutes

17
Testbeds

• Three testbeds UIUC, UCSD, Chiba City

18
Example of experimental results

• Comparison of five policies on UCSD cluster

19
Result evaluation

• An average mean and an average s.d.
• Give a rough valuation on the performance of each
  scheduling policy over a given interval of time
• A relative measure of achievement
• Evaluate how often each run achieves a minimal
  execution time
• The statistical analysis
• Show the significance of the improvement of our
  strategy over other strategies

20
Mean and s.d. of execution time
21
Mean and s.d. summary

• CS achieves 2-7 less execution time than HM and
  HCS
• Using better information prediction results
  better execution time
• CS achieves 2-32 less SD of execution time than
  HM, 1.5-77 less SD than did OSS
• Taking variation information into account results
  in more predictable application behavior

22
Compare metric

• A scheduling policy is better than others if it
  exhibits a lower execution time than another
  policy on a given run
• Five possibilities exist
• best best execution time among five
• good better than three policies but worse than
  one
• averagebetter than two and worse than two
• poor better than one but worse than three
• worstworst execution time of all five

23
Compare results

• Conservative Scheduling is more likely to have a
  best or good execution time than the other
  approaches on all clusters
• Taking account of the average and variation in
  CPU information can significantly improve
  application performance

24
Paired one tailed T-test

• A statistical method used to assess whether the
  means of two groups are significantly different
  from each other
• The result is a set of P-values that indicate the
  possibility that the differences could have
  happened by chance
• a lower P-value means a more significant
  difference between two groups

25
Paired one tailed T-test

• In most cases, P-valueslt0.1 possibility of the
  improvement happening by chance is quite small

26
Conclusion

• Conservative Scheduling policy address both
  heterogeneous and dynamic nature of share systems
• Predictions of expected mean and variance
• Conservative load prediction
• predicted meanpredicted SD
• Better execution time and more predictable
  application behavior

27
Contact

• Lingyun Yang lyang_at_cs.uchicago.edu
• Jennifer M . Schopf jms_at_mcs.anl.gov
• Ian Foster foster_at_mcs.anl.gov
• Paper details
• http//people.cs.uchicago.edu/lyang/work/Schedul
  ing_SC_CameraReady.pdf
• Slides are available from
• http//people.cs.uchicago.edu/lyang/work/Scheduli
  ng.ppt