genetic algorithms for energy efﬁcient virtualized data ... · genetic algorithms for energy...

Genetic Algorithms for Energy Efficient Virtualized Data Centers

Genetic Algorithms for Energy EfficientVirtualized Data Centers

6th International DMTF Academic Alliance Workshop on Systemsand Virtualization Management: Standards and the Cloud

Helmut Hlavacs, Thomas TreutnerUniversity of Vienna, Austria

26.10.2012

1 / 40


Table of Contents

1 Scenario

2 Utilization Trace Data

3 Balanced First Fit

4 Genetic Algorithm

5 Parameters

6 Results

7 Performance Evaluation

8 Conclusions

2 / 40


Abstract

In A Nutshell• Efficiency by dynamic consolidation + workload forecasting• Heterogeneous infrastructure in terms of power, resources• Evaluation of real traces, University of Vienna Central IT Dept.• CPU traces of ≈35 VMs, 4 weeks, 2h resolution, VMware• Business infrastructure scenario: Energy costs are just one of

several parts of operational costs ⇒Use a cost model!• Cost model, configurable penalties for several cost

categories, minimize total weighted costs• Multi-objective combinatorial optimization problem• Comparison of total weighted costs: Balanced First Fit

heuristic, Genetic Algorithm, Load Balancing• Forecasting: (S)ARIMA, Holt-Winters

3 / 40


Scenario

Scenario

Scenario• Highly variable workload intensity, often periodic.• No (little?) number-crunching, its not a HPC cluster etc.• Minimize energy consumption while avoiding

under-provisioning, before reaching 100% utilization!• Queuing issues! Need resources for live migration!• Status costs:

• Energy: Linearly correlated with CPU util, future work: SPECpower• Overloads: Queuing, Bad QoS, loose revenue, non-linear,

ideally continuous function!• Reconfiguration costs:

• Live Migration: Resource intensive process• Server Boots/Shutdowns: Costs energy, puts mechanical/electrical

strain?

4 / 40


Scenario

Non-linear overload cost function

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 0.2 0.4 0.6 0.8 1

Overl

oad C

ost

Utilization

Figure: Markovian M/M/1 queue, P(T > x) = 1 − FT (x) = e−µ(1−ρ)x , ρ asCPU util, service rate µ and max response time x must be supplied

5 / 40


Scenario

Static Over-provisioning

0 0.5 1 1.5 2 2.5 3

Res

ou

rces

Time [Days]

DemandCapacity

6 / 40


Scenario

Peak-provisioning

0 0.5 1 1.5 2 2.5 3

Res

ou

rces

Time [Days]

DemandCapacity

7 / 40


Scenario

Static under-provisioning

0 0.5 1 1.5 2 2.5 3

Res

ou

rces

Time [Days]

DemandCapacity

8 / 40


Scenario

Dynamic Provisioning for Actual Demand

0 0.5 1 1.5 2 2.5 3

Res

ou

rces

Time [Days]

DemandCapacity

9 / 40


Utilization Trace Data

Diagnostic time series plot

40

06

00

80

01

00

0

Hour

!

!

!

! !

!

!

! !

!

!

!

!

!

!

!

!

!

!

!!

!!

!

!

!

!

! !

!

!

!

!

!!

!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!

!

!!

!

!

! !

!

!

!!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!

!

!

!

!!

!

!

!!

!

!

!

!

!

!

!!

!

!

!!

!

!

!

! !!

!!

!

!

!

!

!

!

! !

!

!

!

!

!

!

!!

!

!

!

!!

!

!!

!

!

!!

!

!

!

!

!

!

!

!

!

!

!!

!

!

!

! !

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!

!

!

!

!! !

!

!

!

!

!

!

!

!

!

!

! !

!

!

!

!

!

!!!

!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!!

!

!

!

!

!

!

!

! !

!

!!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

! !

!

!

! !!

!

!

!

!

!

!!

!

!

!

!

! ! !

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!!

!!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

!

9pm 1am 5am 9am 1pm 5pm 10 / 40



Periodic, seasonal resource demand

0

200

400

600

800

1000

1200

1400

1600

1800

2000

2200

0 50 100 150 200 250 300 350

Uti

liza

tio

n [

MH

z]

Interval 11 / 40



Bursty resource demand

0

200

400

600

800

1000

1200

1400

0 50 100 150 200 250 300 350

Uti

liza

tio

n [

MH

z]

Interval

Figure: Diagnostic time series plot of the same trace as in Figure ??.

12 / 40



Complete change in resource demand

50

100

150

200

250

300

350

0 50 100 150 200 250 300 350

Uti

liza

tio

n [

MH

z]

Interval 13 / 40



Seasonal resource demand with a changing mean

40

60

80

100

120

140

160

180

0 50 100 150 200 250 300 350

Uti

liza

tio

n [

MH

z]

Interval

Figure: Diagnostic time series plot of the same trace as in Figure ??.

14 / 40


Balanced First Fit

Balanced First Fit

• Bin packing related heuristic, inherently inflexible• Not influencable by cost model, but evaluated by it• Needs a sorting criteria for the bins, SPECpower ssj2008 score• In a nutshell, three phases:

1 Check servers for utilizations exceeding threshold, if so,remove VMs resource-balanced until not overloaded, add VMs tohomelessVMs

2 Try to map homelessVMs beginning with most energy-efficient.Do this resource-balanced again. If still homelessVMs, force action.

3 Try to consolidate less energy-efficient servers, only if all of itsVMs can be migrated to more efficient servers, andvmConsolidationInertia reached

• Hysteresis control: Turn of servers if rmIdleTimeout reached

15 / 40


Genetic Algorithm

Genetic Algorithm

• Meta-heuristic, directly influencable by cost model• Fitness value is reciprocal to the cost of a solution• Lower cost solutions have higher survival chances• Cross-over, Mutation, Evaluation, Selection• Elitism Selection, Roulette Wheel Selection• Max number of generations, stop if quality not increasing for n

generations ⇒ Ensures good quality and runtime• Multi-threading by using demes, randomly exchanging solutions• Several mutators defined• Solution defined by mapping matrix

• Rows are servers• Columns are VMs

16 / 40


Genetic Algorithm

Genetic Algorithm: Crossover operator

Single point crossover

XFather =

1 0 0 10 1 1 00 0 0 0

;XMother =

0 1 1 00 0 0 11 0 0 0

XSon =

1 0 1 00 1 0 10 0 0 0

;XDaughter =

0 1 0 10 0 1 01 0 0 0

17 / 40


Genetic Algorithm

Genetic Algorithm: swapRM operator

Swap two rows

Xold =

1 0 0 10 1 0 00 0 1 0

Xnew =

0 1 0 01 0 0 10 0 1 0

18 / 40


Genetic Algorithm

Genetic Algorithm: swapVM operator

Swap two columns

Xold =

1 0 0 10 1 0 00 0 1 0

Xnew =

0 0 1 10 1 0 01 0 0 0

19 / 40


Genetic Algorithm

Genetic Algorithm: migrateVM operator

Migrate a VM

Xold =

1 0 0 10 1 0 00 0 1 0

Xnew =

1 1 0 10 0 0 00 0 1 0

20 / 40


Genetic Algorithm

Genetic Algorithm: consolidateRm operator

Move all “1“s to another row

Xold =

1 0 0 10 1 1 00 0 0 0

Xnew =

1 1 1 10 0 0 00 0 0 0

21 / 40


Parameters

Parameters

• 28 days of trace, first 7 reserved for forecasting, 21 for eval• Hundreds of VMs, resampled from the trace data (scaling,

memory alloc)• 26 servers, taken from SPECpower ssj2008, high diversity• Optional linear interpolation to “emulate” more frequent

measurements ⇒ VMware export limitation• Optional forecasting, GNU R, auto-model-building for each VM

in each interval to consider change in workload pattern• (S)ARIMA: Limit parameter search and data, takes very long• Use 95th upper bound as forecast, very conservative!• Non-linear overload costs: For every minute of an interval, for

every VM running on an overloaded host, multiply cost functionvalue with rmUtilizationPenalty and sum up ⇒ Penalizelong intervals, as overloads are longer or harder detectable

22 / 40


Parameters

Relevance Parameter Value

All

cpuUtilizationWarningLevel 0.6memoryUtilizationWarningLevel 0.8utilizationCostFunctionMu 10utilizationCostFunctionAllowedResponseTime 1energyPenalty 600migrationPenalty 1rmUtilizationPenalty 10bootPenalty 1shutdownPenalty 5

Load Balancing variancePenalty 100000

BFF rmIdleTimeoutSeconds 900vmConsolidationInertiaSeconds 600

GA and LB

numberOfThreads 4numberOfGenerations 200maxGenerationsOfFitnessNotIncreased 10sizeOfPopulation 800crossoverRate 0.5exchangeRate 0.1migrateVmRate 0.3swapRmRate 0.1swapVmRate 0.1elitism true

Optional Forecasting requiredPeriodsForForecasting 323 / 40


Parameters

Simulation Input Data

0

200000

400000

600000

800000

1e+06

1.2e+06

1.4e+06

1.6e+06

0 50 100 150 200 250

MH

z

Interval

VM CPU DemandRM CPU Capacity

RM CPU Warning Level Capacity

Figure: The time series of total VM CPU demand, server capacity and quotacapacity within the warning level used in the simulations. 24 / 40


Results

Total weighted costs

0

1e+06

2e+06

3e+06

4e+06

5e+06

6e+06

LB

BF

F

BF

F-H

oW

i

BF

F-A

RIM

A

GA

GA

-HoW

i

GA

-AR

IMA

LB

BF

F

BF

F-H

oW

i

BF

F-A

RIM

A

GA

GA

-HoW

i

GA

-AR

IMA

LB

BF

F

BF

F-H

oW

i

BF

F-A

RIM

A

GA

GA

-HoW

i

GA

-AR

IMA

LB

BF

F

BF

F-H

oW

i

BF

F-A

RIM

A

GA

GA

-HoW

i

GA

-AR

IMA

Co

st

EnergyMigration

Boot

ShutdownOverload

900s1800s3600s7200s

25 / 40


Results

Dynamic Provisioning

0 0.5 1 1.5 2 2.5 3

Res

ou

rces

Time [Days]

DemandCapacity

26 / 40


Results

-30

-20

-10

0

10

20

30

40

50

60

0 50 100 150 200 250 300 350 400 450 500

Ov

erp

rov

isio

nin

g [

%]

Interval

(Available-Used)/Used

Figure: Provisioning efficiency for an interval length of 3600 s and theBFF heuristic without forecasting.

27 / 40


Results

0

20

40

60

80

100

120

140

0 50 100 150 200 250 300 350 400 450 500

Ov

erp

rov

isio

nin

g [

%]

Interval


Figure: Provisioning efficiency for an interval length of 3600 s and theBFF heuristic with Holt-Winters forecasting.

28 / 40


Results

0

5

10

15

20

25

30

35

40

45

50

55

0 200 400 600 800 1000 1200 1400 1600 1800 2000

Ov

erp

rov

isio

nin

g [

%]

Interval


Figure: Provisioning efficiency for an interval length of 900 s and the BFFheuristic with Holt-Winters forecasting.

29 / 40


Results

Changing the overload penalty

0

100000

200000

300000

400000

500000

BF

F-H

oW

iG

A-H

oW

i

BF

F-H

oW

iG

A-H

oW

i

BF

F-H

oW

iG

A-H

oW

i

BF

F-H

oW

iG

A-H

oW

i

BF

F-H

oW

iG

A-H

oW

i

BF

F-H

oW

iG

A-H

oW

i

BF

F-H

oW

iG

A-H

oW

i

Co

st

Overload Cost

EnergyMigration

BootShutdown

Overload

1000500200100402010

30 / 40


Results

Changing the migration penalty

0 100000 200000 300000 400000 500000 600000 700000 800000 900000

BF

F-H

oW

i

GA

-HoW

i

BF

F-H

oW

i

GA

-HoW

i

BF

F-H

oW

i

GA

-HoW

i

Co

st

Migration Cost

EnergyMigration

Boot

ShutdownOverload

100501

31 / 40


Results

Changing the energy penalty

0

200000

400000

600000

800000

1e+06

1.2e+06

1.4e+06

BF

F-H

oW

i

GA

-HoW

i

BF

F-H

oW

i

GA

-HoW

i

BF

F-H

oW

i

GA

-HoW

i

Co

st

Energy Cost

EnergyMigration

Boot

ShutdownOverload

24001200600

32 / 40


Results

VM consolidation inertia, Holt-Winters forecasting

0

100000

200000

300000

400000

500000

600000

700000

800000

900000

1e+06

0 300 600 900 1800 3600 7200

Co

st

VM Consolidation Inertia [s]

EnergyMigration

BootShutdown

Overload

33 / 40


Results

VM consolidation inertia, no forecasting

0

200000

400000

600000

800000

1e+06

1.2e+06

0 300 600 900 1800 3600 7200 14400 28800

Co

st

VM Consolidation Inertia [s]

EnergyMigration

BootShutdown

Overload

34 / 40


Performance Evaluation

Performance: Hardware Platform Specification

Platform: Low Power Desktop ServerCPU AMD E-350 PhenomII X4 955 Intel Xeon E5-2670

CPU Frequency 1.6 GHz 3.2 GHz 2.6 GHzCPU Cores 2 4 8

CPU L2-Cache 2x512 KiB 4x512 KiB 8x256 KiBCPU L3-Cache N/A 6 MiB shared 20 MiB shared

CPU TDP 18 W 125 W 115 WMemory 2 GiB 16 GiB 64 GiB

Table: Description of platforms used in the performance evaluations.

35 / 40



Balanced First Fit

0

0.5

1

1.5

2

2.5

BF

F

BF

F-H

oW

i

BF

F

BF

F-H

oW

i

Wal

ltim

e [s

]

AMD Phenom II X4AMD E-350

Figure: Runtime per interval of BFF with/without Holt-Winters forecastingon a low power CPU and a high-end desktop CPU.

36 / 40



Genetic Algorithm

0

50

100

150

200

250

300

350

GA

GA

-Ho

Wi

GA

GA

-Ho

Wi

Wal

ltim

e [s

]

AMD Phenom II X4AMD E-350

Figure: Runtime per interval of GA with/without Holt-Winters forecastingon a low power CPU and a high-end desktop CPU.

37 / 40



Genetic Algorithm Parallelization Speedup

0

0.5

1

1.5

2

2.5

2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

Spee

dup

Numer of Cores

AMD PhenomII X4 3.2GHzDual Xeon E5-2670 2.6GHz

Figure: Speedup achieved by parallelizing the genetic algorithm.

38 / 40


Conclusions

Conclusions

• Flexible cost model feeding into a GAs fitness function• Easy adaptation to diverse optimization demands• Case study parameter sets, drastic reduction of total costs• For long intervals, forecasting is essential• Heuristics faster, but inflexible to changing parameters

(energy price, overload costs etc.)• GA can do load balancing by changing single parameter• Future Work:

• Non-linear, heterogeneous live migration costs• Heterogeneous VM overload costs (priorities)• Penalizing co-existence of VM pairs on a host (customer isolation,

performance issues, security)• Speed up GA by storing final solution of the last n intervals,

replaying them to solution population• GA multi-threading speedups?!

39 / 40


Conclusions

Q&A

Thank you for your attention!

http://www.metavisor.org was sponsored by the Internet FoundationAustria within the Netidee 2012 funding programme -

http://www.netidee.at

40 / 40

genetic algorithms for energy efﬁcient virtualized data ... · genetic algorithms for energy...

Documents