Advanced Energy Management

in Cloud Computing multi data center environments

Giuliana Carello, DEI, Politecnico di Milanocarello@elet.polimi.it

Danilo Ardagna, Antonio CaponePolitenico di Milano, 12 June 2012

CLOUD CONSUMPTION

About 0.5% of global electric power consumption is due to Data Centers (DC)

In developed country: ‐UK: 2.2-3.3% ‐USA: 1.5%

From the environmental point of view:‐2% of global CO2 emissions

2005

2020

0.0% 5.0% 10.0% 15.0% 20.0% 25.0%

% European IT consumption

Cellular phone NetworkTelecom NetworkServer & Data Center

Source: EU Commission

ENVIRONMENTAL IMPACT

High consumption, environmental impact and energy costs of Cloud Computing

A way out: “GREEN” CLOUD COMPUTING

New Servers costsEnergy and cooling costs

IT costsNumber of Servers (M units)

…and costsEnvironmental impact

Source: EU Commission

OUR APPROACH

Green Cloud project: Innovative strategies for resource allocation in Cloud Computing multi data centers

systems Exploiting different energy costs and different workload due to geographically

distributed data centers Integrated Data Center and network management Green energy sources utilization

Definition of Linear Programming Optimization models Scenario definition and analysis

SCENARIO

Optimization over 24 hours for each time band Set of geographically distributed DC Traffic profile for each Data Center over the day Full-connected network Possibility to redirect requests from one DC to another Energy cost for both DC and network: fixed + linear

NETWORK

Request forwarding is due to: Lower energy cost Tradeoff between DC and network costs Capacity limits not reached

BROWN MODEL

DECISION VARIABLES:

Requests rate of type k incoming to DC i served in DC j with a type l server

OBJECTIVE FUNCTION:

Minimization of the costs over 24 hours Cost for operating servers Cost for switching on/off servers Bandwidth utilization Number of active link and link switching costs

BROWN MODEL

CONSTRAINTS:DATA CENTERS

The whole incoming traffic has to be served Each request must be served by suitable server DCs can handle a finite number of requests Server have a utilization limit

BROWN MODEL

CONSTRAINTS:DATA CENTERS

The whole incoming traffic has to be served Each request must be served by suitable server DCs can handle a finite number of requests Server have a utilization limit

NETWORKS Links have limited bandwidth

BROWN MODEL

CONSTRAINTS:DATA CENTERS

The whole incoming traffic has to be served Each request must be served by suitable server DCs can handle a finite number of requests Server have a utilization limit

NETWORKS Links have limited bandwidthSWITCHING

Implemented in AMPL language using IBM ILOG CPLEX 12.1 solver

GREEN MODEL

ASSUMPTIONS: DCs can be powered through renewable sources:‐ Solar‐ Geothermic‐ Wind

Limited amount of green energy available Only autonomous production Minor cost with respect to brown energy sources

IMPLICATIONS: Green powered DCs are preferred CO2 reduction

DATA CENTER MAP

Data Center locations

INCOMING TRAFFIC

Starting from the basic profile: Re-scaling according to the geographical location Temporal shift to consider time zone differences

Analysis of the model behavior with respect to a growing number of incoming requests:2 – 3 – 16 – 30 – 40 billions of daily requests

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 240

1

Time band

ENERGY COST

Differential trend according to: Energy market studies Geographical location Time band

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 240

10

20

30

40

50

60

70

West USA

East USA

Europe

Asia

Time band

€/M

Wh

GREEN DATA CENTERS

Location derived from green energy sources availability

GREEN ENERGY AVAILABILITY

Different trends and energy production according to the different green energy sources

Analysis based on the dimension and location of the DCs

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 240

50

100

150

200

250

Mountain View, CA - Solar power

Seattle, WA - Geothermal power

Frankfurt, GER - Wind power

Sao Paolo, BRA - Solar power

Time bands

kWh

REQUESTS REDIRECTION

Requests are redirected in the macro-area where the DC energy cost is lower

0

50

100

150

200

250

300

350

05

10152025303540

0

100

200

300

400

500

600

05

1015202530354045

0

20

40

60

80

100

120

140

0

10

20

30

40

50

60

70

050

100150200250300350400450500

0

10

20

30

40

50

60

70

West US East US Europe Asia

Activ

e se

rver

s€/

MW

h

POLICIES TO BE COMPARED

Brown Model: Possibility of request redirection from one DC

to another

Green Model: Extension of the Brown Model taking into

account green energy sources

Base Case: Requests execution in local

EXPENSES REDUCTION

2 3 16 30 400

2000

4000

6000

8000

10000

12000

14000

16000

35% 39%

39%

37%

Base

Brown

Billions of daily requests

Ener

gy e

xpen

ses

(€/d

ay)

Unf

easi

ble

solu

tion

BROWN MODEL

2 3 16 30 400

50,000

100,000

150,000

200,000

250,000

300,000

350,000

400,000

450,000

500,000

12% 7%

9%

10%

BaseBrown

Billions of daily requests

Ener

gy c

onsu

mpti

on (k

Wh/

day)

Unf

easi

ble

solu

tion

ENERGY CONSUMPTION

Model priority: minimizing expenses Higher energy consumption due to network utilization for request forwarding

2 3 16 30 400

2000

4000

6000

8000

10000

12000

14000

56% 49%

43%

40%

Base

Green

Billions of daily requests

Ener

gy e

xpen

ses

(€/d

ay)

Unf

easi

ble

solu

tion

EXPENSES REDUCTION

GREEN MODEL

ENERGY CONSUMPTION

2 3 16 30 40 -

100,000

200,000

300,000

400,000

500,000

600,000

DC Green

DC Brown

Network

Billions of daily requests

Ener

gy c

onsu

mpti

on (k

Wh)

Green energy saturation Possibility to better exploit the requests redirection by investing in

green energy sources

CO2 EMISSIONS

2 3 16 30 400

50

100

150

200

250

300

350

400

57% 41%

11%

8%

7%BrownGreen

Billions of daily requests

CO2

emis

sion

s (t

ons/

day)

Comparison of brown and green models based on the CO2 emissions Maximum amount of green energy availability fixed at 8%

MODEL FOOTPRINT Reduction of greenhouse gases according to

different green energy availability

CO2 EMISSIONS

MODEL GAP

Two hours time limit, average gap 5%

MODEL SOLVING TIME

FUTURE DEVELOPMENT

The model proved to be: Robust, flexible and scalable

POSSIBLE EXTENSIONS: Non fully connected network topology SLA considerations and response time Tests on larger instances

CONCLUSIONS

ECONOMICAL POINT OF VIEW: Optimization of Network and DCs jointly Expenses reduction up to 40%

ENVIRONMENTAL POINT OF VIEW: Reduction of greenhouse gas emissions Optimizing costs linked to the Carbon Credit system Promoting the use of green energy sources

THANK YOU FOR THE

ATTENTION!!!

ANY QUESTIONS?