Energy Saving in Cloud Data Centers

Energy Saving in Cloud Data Centers


Presentation of the results of EU FP7 FIT4Green

Corentin Dupont

cdupont@create
-
net.org


Presentation


FIT4
Green

seeks energy saving policies for DCs, enhancing the
effects inside a federation by an aggressive strategy for
reducing the energy consumption in ICT


We aim at reducing cost for companies


Strengthening
competitive position


FIT4
Green
needs to be DC framework agnostic:


Demonstrated in Cloud computing, Traditional computing, Super
computing and Networking

Table of Contents


Introduction

&

overview


Requirements


Optimizer

design


SLA

Constraints


Power

Objective

Model


Heuristics


Experiments

on

Cloud

Test
-
bed


Scalability

Evaluation


Conclusion

&

Future

work





By shifting activities that do not
interfere with each other to the
same room and switch off
unused electrical consumers in
the other rooms




By moving activities to rooms
with more energy efficient light
bulbs or other more efficient
electrical devices that suffice
the needs

Common sense: saving energy at home

Page
4


Energy Saving Strategies

Page
5


VMs are consolidated and
unused servers are turned off





VMs are allocated to “more
efficient” servers/data
centres
:
incremental “cost” considered.



Energy Savings in a Federation

Page
6

PUE: 1,3

PUE: 1,2

PUE: 1,4

PUE: 2,4

PUE: 2,1

PUE: 1,8

PUE: 1,9



takes

the

difference

in

PUEs

into

account,

and

where

applicable

reallocate

VMs

to

the

Cloud

DC

with

a

better

efficiency
.


In

addition

a

larger

resource

pool

provides

larger

optimization

opportunities

(smaller

local

resource

buffers)
.


Moreover,

CUE

differences

drive

emission

optimizations


Federation
-
Enabled Optimizations

“Federation” policies
seek to:



Relocate VMs to capitalize on
geographical characteristics like:


Season &Temperature differences


Time
-
zone differences
–

‘Follow the
Sun’


Energy source differences



Relocate

VMs to capitalize on data
centres

characteristics like:


Equipment & Infrastructure
differences


PUE & CUE differences


Cogeneration options

Page
7

All strategies are ranked through their Energy KPIs.

Cloud Load

Page
8

Load

Time


Tetris vs.


Page
9




In

High

Performance

Computing

DCs
,

the

order

of

job

execution

is

optimized




In

Cloud

Computing

DCs
,

the

allocation

of

new

workload

is

being

considered




In

Traditional

Computing

DCs
,

workload

is

being

reallocated

based

on

energy/CO
2

efficiency
.


High Level View

Página
10


DC Federation


Model

Optimization

Reconfiguration

Monitoring

Power Calculator

Data Centre Monitoring and Automation Framework

Power Consumption Predictions

Updated Information

List of Suggested Actions

Energy Model

Page
11

Queue


Software
Applications

Network Topology

Framework

Capabilities

Storage

Server



ICT Resources

Queue


Software
Applications

Network Topology

Framework

Capabilities

Storage

Server



ICT Resources












n
i
i
core
idle
CPU
n
i
i
i
idle
CPU
CPU
t
E
t
E
t
CL
f
V
t
P
t
E
1
_
_
1
2
_
)
(
)
(
Optimizer

All strategies are ranked through their Energy KPIs

Metamodel

Topology

SLAs

Policies

Automatic

Rule

Automatic

Rule

Automatic

Rule

Automatic

Rule

Constraint

Constraint

Constraint

Constraint

Constraint Programming

Optimization Engine with Energy/Emission Goal

ACTIONS

Requirements

Abstracting out the constraints

•

Flexibility, extensibility

•

Deep

exploration of the
search

space

Framework Design

SLA CONSTRAINTS

SLA constraints flow

SLA CONSTRAINTS

SLA constraints examples

Category

Constraint

Approach

LoC

Hardware

HDD

Choco + ext. Entropy

121+(25)

CPUCores

Entropy (‘fence’)

0+(25)

CPUFreq

Entropy (‘fence’)

0+(25)

RAM

Choco + ext. Entropy

123+(25)

GPUCores

Entropy (‘fence’)

0+(25)

GPUFreq

Entropy (‘fence’)

0+(47)

RAIDLevel

Entropy (‘fence’)

0+(47)

QoS

MaxCPULoad

Choco

+ ext. Entropy

90+(25)

MaxVLoadPerCore

Choco

+ ext. Entropy

109+(25)

MaxVCPUPerCore

Choco

+ ext. Entropy

124+(25)

Bandwidth

Entropy (‘fence’)

0+(49)

MaxVMperServer

Entropy (‘capacity’)

0+(25)

Availability

PlannedOutages

Choco

+ ext. Entropy

Future Work

Availability

Choco

+ ext. Entropy

Future Work

Additional

Metrics

Dedicated

Server

Entropy (‘capacity’)

0 + (25)

Access

Entropy (‘fence’)

0 + (25)

POWER OBJECTIVE MODEL

Total
Reconf
.
Energy

Total Instant. Power

Energy

Migrations

Energy

On/Off

Power Servers Idle

Power VMs

Power Network

*
Reconf

Time

Power
Calculator

HEURISTICS

Root node:

no VM is allocated

First level node:


VM1 allocated on S1

First level node:


VM2 allocated on S1

First
level

node
:


VMx

allocated

on
Sy

Leaf

node
:

all
VMs

are
allocated

Leaf

node
:

all
VMs

are
allocated

Leaf

node
:

all
VMs

are
allocated

Leaf

node
:

all
VMs

are
allocated

At

each

level
: call F4G
branching

heuristic
. If a
constraint

is

broken
,
backtrack

to go up.

At

leaf

level
: note down
the solution and the
energy

saved
,
then

backtrack

to
find

a
better

solution.

First
level

node
:


VMx

allocated

on
Sy

First
level

node
:


VMx

allocated

on
Sy

Heuristics

Select VM on the least
energy

efficient server and
least
loaded

server

Call the F4G VM
selector

VM
selected

Select Server
which

is

the
most

energy

efficient server
and
most

loaded

server

Call the F4G
Server
selector

Server
selected

Select Server
which

is

empty

and the least
energy

efficient server

Call the F4G Server
selector

Server
selected

Composable
heuristics

•

Candidate VM

for migration

•

Target server

for migration

•

Candidate Server
for extinction

Heuritics

To
sum

up…

Experiments on Cloud
Testbed

Node
Controller
Node
Controller
Node
Controller
Node
Controller
Node
Controller
Node
Controller
Node
Controller
Cluster
Controller
Power and
Monitoring
Collector
Cluster
Controller
Cloud
Controller
Task scheduler
FIT
4
Green VMs
Blade Enclosure
1
Blade Enclosure
2
Lab

trial ressources

Enclosure 1

Enclosure 2

Processor

model

Intel Xeon
E5520

Intel Xeon
E5540

CPU

frequency

2.27GHz

2.53GHz

Cpu
&

Cores

Dual cpu
–

Quad
core

Dual cpu
–

Quad
core

RAM

24 GB

24GB

•

DC1: 4 BL 460c blades using
VMWare

ESX v4.0 native hypervisor, 3
blades for Cluster and Cloud Control

•

DC2: 3 BL460c blades using
VMWare

ESX v4.0 native hypervisor, 2
blades for Cluster Control and Power and Monitoring System.

Experiments on Cloud
Testbed

Total number of active virtual machines during full week of work

Number of
active VMs

Time

Active SLAs constraints:

•

Max
vCPU

per core = 2

•

Min VM Slot = 3

•

Max VM Slot = 6

Lab

trial
Workload

Experiments on Cloud
Testbed

Final test results for the various configurations

Configuration

Data
Centre 1

Data
Centre 2

Energy for
Federation

Without

FIT
4
Green

6350 Wh

4701 Wh

11051 Wh

With

FIT
4
Green

Static

Allocation

5190 Wh

4009 Wh

9199
Wh

Saving
16.7%

With

FIT
4
Green

Dynamic

Allocation

5068 Wh

3933 Wh

9001 Wh

Saving
18.5%

With

FIT
4
Green

Optimized

Policies

4860 Wh

3785 Wh

8645
Wh

Saving
21.7%

CONCLUSION & FUTURE WORK


Energy

aware

resource

allocation

in

datacenters


Flexibility

&

extensibility


Saves

up

to

18
%

in

HP

experiment


Scalability

with

parallel

processing



Future

work
:


SLA

re
-
negotiation


Green

SLAs


2
nd

phase tests numeric results

Numeric

results

Single

site tests

Federated

site tests

Traditional
DC
testbed

Around 30%

From 28%

to
48%

Supercomputing
DC

testbed

From 4% to

28%

From 30%

to 42%

Cloud computing
DC

testbed

From 10% to

24%

From 17% to 21%

Page
26

Scalability Evaluation

#

Configuration

Placement constraints activated

1

1

datacenter

none

2

1 datacenter

with

overbooking factor=2

“
MaxVCPUPerCore
”

constraint

set

on

each

server

3

2

federated

datacenters

“Fence”

捯nst牡rnt

s整

潮

敡捨

噍

FIT4
Green
Plug
-
in

Page
28

Single

allocation

Find

the

most

energy

efficient

and

suitable

resource

for

a

new

Workload
.



Global

optimization

Rearrange

the

resources

in

a

way

that

saves

maximum

amount

of

energy

or

carbon

emission
.

Page
29

Optimizer

entry points