Resource management in IaaS cloud · 2017-01-30 · VM5 6 min(#onlineNodes) = 3 sol #1: 1m,1m,2m. dynamic schedulers quality at a price N3 cpu mem N1 cpu mem N4 cpu mem N2 cpu mem

Fabi

en H

erm

enier

Resource management in IaaS cloud

where to place the VMs

how much to allocate?

1What can a resource manager do for you

3

2Models of VM scheduler

4

3Implementing VM schedulers

5

1What can a resource manager do for you

6

central piece of code

provide the awaited Quality of Service address the management objectives

?

the RM should solve that

VM1VM2 V

M3

VM4 VM5

VM

6

N1 N2

N4N3

cpu cpu

cpucpu

mem

mem

mem

mem

the RM should prevent that

VM1

VM

3

VM4VM5

N1 N2

N4N3

cpu cpu

cpucpu

mem

mem

mem

mem VM6

VM2

a contract where a service is formally defined

Service

Agreement

Levelperformance indicator

MTTRMTBF

availability (MTBF/MTTR + MTBF)pricing

penalties…

uptime as the only Key Performance Indicator (KPI)?

Cloud Performance SLAs

What customers wantguaranteed capacities (CPU, memory, bw)

guaranteed latencies/throughput (network, load balancer, disk)

What providers give-

the objectiveprovider side

min(x) or max(x)

min(penalties)min(Total Cost Ownership)

min(unbalance)

atomic objectives

…

min(αx + β y)

composite objectivesusing weights

useful to model sth. you don’t understand ?How to estimate coefficients ?

min(α TCO + β VIOLATIONS)max(REVENUES)

€ as a common quantifier:

?Should a client ask for an objective

?Should a client ask for an objective

no if the provider cannot prove it is achieved

proving a latency is minimal …one can check the latency is below a threshold

VM-host affinity (DRS 4.1)

Dedicated instances (EC2)

MaxVMsPerServer (DRS 5.1)

apr. 2011

mar. 2011

sep. 2012

The constraint needed in 2012

?? 2013

VM-VM affinity (DRS)

2010 ?

Dynamic Power Management (DRS 3.1)

2009 ?

SLOs and objectives are evolving

there is no holy grail

2Models of VM schedulers

20

monitoring data

VM queue

actu

ator

sVM scheduler

cloud model

decisions

sche

dulin

g m

odel

s

static dynamic schedulers

static dynamic resource allocation

VM1

VM2 VM

3

VM4VM5

VM

6

N1 N2

N4N3

cpu cpu

cpucpu

mem

mem

mem

mem

static schedulers

VM7

VM1

VM2 VM

3

VM4VM5

VM

6

N1 N2

N4N3

cpu cpu

cpucpu

mem

mem

mem

mem

static schedulers

VM7

VM1

VM2 VM

3

VM4VM5

VM

6

N1 N2

N4N3

cpu cpu

cpucpu

mem

mem

mem

mem

static schedulers

VM7

Combinatorial problem !

past decisions impact the future ones

VM1

VM2 VM

3

VM4VM5

VM

6

N1 N2

N4N3

cpu cpu

cpucpu

mem

mem

mem

mem

static schedulers

buffering might help

VM1

VM2 VM

3

VM4VM5

VM

6

N1 N2

N4N3

cpu cpu

cpucpu

mem

mem

mem

mem

static schedulers


VM1

VM2 VM

3

VM4VM5

VM

6

N1 N2

N4N3

cpu cpu

cpucpu

mem

mem

mem

mem

static schedulers

VM7


VM1

VM2 VM

3

VM4VM5

VM

6

N1 N2

N4N3

cpu cpu

cpucpu

mem

mem

mem

mem

static schedulersVM7


VM1

VM2 VM

3

VM4VM5

VM

6

N1 N2

N4N3

cpu cpu

cpucpu

mem

mem

mem

mem

static schedulersVM7


/!\ incoming ratestarvation

static schedulersmanipulate the VM queue only

react on VM arrival

static schedulersproseasy to model simple standard actions manage a few elements

static schedulerscons

hard to perform fine grain optimisation

VM1

VM2 VM

3

VM4VM5

VM

6

N1 N2

N4N3

cpu cpu

cpucpu

mem

mem

mem

mem

dynamic schedulers

consider every VMs

VM1

VM2 VM

3

VM4VM5

VM

6

N1 N2

N4N3

cpu cpu

cpucpu

mem

mem

mem

mem

dynamic schedulers

consider every VMs

VM1

VM2 VM

3

VM4VM5

VM

6

N1 N2

N4N3

cpu cpu

cpucpu

mem

mem

mem

mem

dynamic schedulers

consider every VMs

VM7

VM1

VM2 VM

3

VM4VM5

VM

6

N1 N2

N4N3

cpu cpu

cpucpu

mem

mem

mem

mem

dynamic schedulers

consider every VMs

VM7

VM1

VM2 VM

3

VM4VM5

VM

6

N1 N2

N4N3

cpu cpu

cpucpu

mem

mem

mem

mem

dynamic schedulers

consider every VMsVM7

dynamic schedulersmanipulate the VM queue

reconfigure the schedule with migrations

dynamic schedulersa migration takes time

dynamic schedulersa migration impacts performance

dynamic schedulers

N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

?

dependency management

VM1

VM1

VM2 VM2

VM3

VM4

VM5VM5

VM

6

VM

6

VM3

VM4

dynamic schedulersdependency management

N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1 VM2

VM3

VM4

VM5V

M6


N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1 VM2

VM3

VM4

VM5V

M6


N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1 VM2

VM3

VM4

VM5V

M6


N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1

VM2

VM3

VM4

VM5V

M6

dynamic schedulerscyclic dependencies

N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

VM1

VM5VM3

VM4

anti-affinity(VM3,VM4)min(#onlineNodes)

?

N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

VM1

VM4VM3

VM5

anti-affinity(VM3,VM4)min(#onlineNodes)


N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

VM1

VM5VM3

VM4

a pivot to break the cycle

fix or prevent the situation ?


N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

VM1

VM5VM3

VM4a pivot to break the cycle



N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

VM1VM5

VM3

VM4a pivot to break the cycle



N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

VM1VM5

VM3 VM4

a pivot to break the cycle


dynamic schedulersquality at a price

N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1 VM2

VM3

VM4

VM5

VM

6

min(#onlineNodes) = 3


N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1 VM2

VM3

VM4

VM5

VM

6



N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1 VM2

VM3

VM4

VM5

VM

6



N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1

VM2

VM3

VM4

VM5

VM

6



N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1

VM2

VM3

VM4

VM5

VM

6


sol #1: 1m,1m,2m


N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1 VM2

VM3

VM4

VM5

VM

6


sol #1: 1m,1m,2m


N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1 VM2

VM3

VM4

VM5

VM

6


sol #1: 1m,1m,2m


N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1 VM2

VM3

VM4

VM5

VM

6


sol #1: 1m,1m,2m


N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1 VM2

VM3

VM4

VM5

VM

6


sol #1: 1m,1m,2m

sol #2: 1m,1m1m


N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1 VM2

VM3

VM4

VM5

VM

6


sol #1: 1m,1m,2m

sol #2: 1m,1m1m

lower MTTR (faster)


the objective should reflect reconfiguration costs

min(MTTR), min(#migrations),…

dynamic schedulersproscontinuous optimisation through reconfiguration

dynamic schedulerscons

harder to model harder to scale

technically expensive costly reconfiguration

benefits depends on

dynamic scheduling is rare in public or large clouds

dynamic scheduling for the win

in theory yes but that’s theory

the workload the objective/ SLAs the infrastructure

?

static dynamic

schedulersstatic dynamic

resource allocation

sche

dulin

g m

odel

s

static resource allocation

cpu, ram, i/o, bandwidth allocated once for all

allocation != utilisation

N1

mem

no sharing conservative allocation

pcpu

VM1

VM2

VM3

N1 pcpu

mem

sharing (overbooking)

performance loss with concurrent accesses

acceptable if stated in the SLA

static resource allocation

VM1

VM2

VM3

dynamic resource allocation

N1 pcpu

mem

cpu, ram, i/o, bandwidth allocated can be revised

N1 pcpu

mem

to fix violations

VM1

VM3

VM2


N1 pcpu

mem


N1 pcpu

mem

to fix violations

VM1

VM2


N1 pcpu

mem


N1 pcpu

mem

to fix violations

VM1

VM2


N1 pcpu

mem

N1 pcpu

mem

to support vertical elasticity

VM1

VM2

VM3


N1 pcpu

mem

N1 pcpu

mem

VM1

VM2


VM3


N1 pcpu

mem

N1 pcpu

mem

VM1

VM2



N1 pcpu

mem

N1 pcpu

mem

VM1

VM2


VM2

benefits depends on

dynamic resource allocation for the win

common on CPU + overbooking (inc. hosting capacity)

the workload the objective/ SLAs the infrastructure

?exceptional for memory

(huge performance loss)

RECAP50

The VM scheduler makes cloud benefits real

51

no holy grail

52

think about what is costly

53

static scheduling for a peaceful life

54

dynamic scheduling to cease the day

55

with great power comes great responsibility

56

Coding a VM scheduler3

static or dynamic scheduler allocation

does the workload/SLA/objective requires migration

VM scheduling is hard

?

?

maximum duration to schedule ?

issues with large infrastructures or hard problems

fast adhoc heuristics despite corner cases

some use biased complete approaches (linear programming, constraint programming)

VM scheduling is NP-Hard

like him

vector packing problemitems with a finite volume to

place inside finite bins

the basic to model the infra. 1 dimension = 1 resource

a generalisation of the bin packing problem

VM1

VM

3

N1 cpu

mem

VM2

VM

4

N2 cpu

mem

?Which resource can be modeled as a

packing dimension

?Which resource can be modeled as a

packing dimension

CPU, memory, disk IO, licences cardinality, network boundaries, …end to end network

how to support migrations

temporary, resources are used on the source and the destination nodes

how to support migrationsa simple way

n-phases vector packing

VM duplication between 2 phases to simulate the migration

hard to manipulate to compute long term previsionsN3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1 VM2

VM3

VM4

VM5

VM

6

easy to implement





mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1 VM2

VM3

VM4

VM5

VM

6

easy to implementVM4





mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1 VM2

VM3 VM5

VM

6






mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1 VM2

VM3 VM5

VM

6

easy to implement

VM

6

VM4





mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1 VM2

VM3 VM5


VM

6





mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1 VM2

VM3 VM5

easy to implementVM1VM4

VM

6





mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM2

VM3 VM5


VM

6

VM1

how to support migrationsthe alternative way

1-phase

+ compute dependencies step by steps

N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1 VM2

VM3

VM4

VM5

VM

6

ok to implement

N5 cpu

memVM7

offline(N2) + no CPU sharing

cycles #migrations/!\


N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1 VM2VM3VM5V

M6

N5 cpu

mem

VM7 VM4


N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2 cpu

mem

VM1

VM2

VM3

VM5V

M6

N5 cpu

mem

VM7

1) migrate VM2, migrate VM4, migrate VM5

VM4


N3 cpu

mem

N1 cpu

mem

N4 cpu

mem

N2

VM1

VM2

VM3

VM5V

M6

N5 cpu

mem

VM7

1) migrate VM2, migrate VM4, migrate VM52) shutdown(N2), migrate VM7

VM4

coarse grain staging delay actions

mig(VM2)

mig(VM4)

mig(VM5)

off(N2)

mig(VM7)

timestage1 stage2

the clean way to model space and time

Resource-Constrained Project Scheduling Problem

N1

N2

N3

N4

time

VM1

VM5VM6

VM3VM7

VM4VM2 off

0

VM1

VM3

VM7

N5

VM5

VM4VM2

VM6

3 4 8



N1

N2

N3

N4

time

VM1

VM5VM6

VM3VM7

VM4VM2 off

0

VM1

VM3

VM7

N5

VM5

VM4VM2

VM6

3 4 8



N1

N2

N3

N4

time

VM1

VM5VM6

VM3VM7

VM4VM2 off

0

VM1

VM3

VM7

N5

VM5

VM4VM2

VM6

3 4 8



N1

N2

N3

N4

time

VM1

VM5VM6

VM3VM7

VM4VM2 off

0

VM1

VM3

VM7

N5

VM5

VM4VM2

VM6

3 4 8



N1

N2

N3

N4

time

VM1

VM5VM6

VM3VM7

VM4VM2 off

0

VM1

VM3

VM7

N5

VM5

VM4VM2

VM6

3 4 8



N1

N2

N3

N4

time

VM1

VM5VM6

VM3VM7

VM4VM2 off

0

VM1

VM3

VM7

N5

VM5

VM4VM2

VM6

3 4 8

the pure way to support migrations


1 resource per (node x dimension), bounded capacity

tasks to model the VM lifecycle. height to model a consumption width to model a duration

at any moment, the cumulative task consumption on a resource cannot exceed its capacity

comfortable to express continuous optimisation

very hard to implement (properly)

From a theoretical schedule to a practical one

duration may be longer

0:3 - migrate VM4 0:3 - migrate VM5 0:4 - migrate VM2 3:8 - migrate VM7 4:8 - shutdown(N2)

convert to an event based schedule

- : migrate VM4 - : migrate VM5 - : migrate VM2 !migrate(VM2) & !migrate(VM4): shutdown(N2) !migrate(VM5): migrate VM7

Back to vector packing

based approaches

Fit Fit Decrease (FFD)basic VM scheduling

sort VMs in desc order for each VM pick the first suitable node



difficult VMs first !



difficult VMs first !enough free resources

N1 N2 N3 N4

V1

V2

V3

V4

Fit Fit Decrease (FFD)example

N1 N2 N3 N4

V1

V2

V3

V4


N1 N2 N3 N4

V1

V2

V3

V4


N1 N2 N3 N4

V1

V2

V3

V4


N1 N2 N3 N4

V1

V2

V3

V4


multi-dimension sorting ?

V1

V2

V3

V4 easy, 1 dimension is varying

V1

V3

V4 easy, uniform variation

?V1

V3

V4

multi-dimension sorting ?

V1

V2

V3

V4 easy, 1 dimension is varying

V1

V3

V4 easy, uniform variation

V1

V3

V4 sort by dimension

aggregate dimensions find the most critical one …

Balancing VMs

Why?

Why?to reduce loss in terms of failures to reduce hotspots to absorb load spikes

balancingmin(stddev([load(n ), … , load(n )]))

in theory

1 i

min(max([load(n ), … , load(n )]),min([load(n ), … , load(n )]),)

1 i

1 i

OR

Worst Fit Decrease (WFD)balancing, a practice

sort VMs in desc order for each VM pick the suitable node with the highest remaining space

balancing over multiple dimensions

?

dimension normalisation worst dimension dimension aggregation current state vs. next state

balancing over multiple dimensions

?

VM-host affinity (DRS 4.1)

Dedicated instances (EC2)

MaxVMsPerServer (DRS 5.1)

apr. 2011

mar. 2011

sep. 2012

The constraint needed in 2012

?? 2013

VM-VM affinity (DRS)

2010 ?

Dynamic Power Management (DRS 3.1)

2009 ?

What about the SLOs ?

SLOs inject additional rules

placement constraint (anti)affinity rules

VM-VM, VM-PM, relative, absolute

temporal constraint precedences, parralelism, sequence

fault tolerance, hw. compatibility, security

control between actionscontrol, performance

counting constraint

per node, group of nodes, resourceperformance, licensing restriction

… …

discrete constraints

spread(VM[1,2]) ban(VM1, N1) ban(VM2, N2)

N1

N2

N3

VM1

VM2

N1

N2

N3

VM1

VM2

continuous constraints

>>spread(VM[1,2]) ban(VM1, N1) ban(VM2, N2)

harder scheduling problem (think about actions interleaving)

“simple” spatial problem

soft constraints

hard constraints

must be satisfied all or nothing approach not always meaningful

satisfiable or not internal or external penalty model

harder to implement/scale hard to standardise ?

spread(VM[1..50])

mostlySpread(VM[1..50], 4, 6)

energy efficient VM scheduler

(Environmental Protection Agency, 2007)

of the 2005 budget

1.5%

in 2010 ?3 %USA

1Consume less2003

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●

●

●●●

●

●●●●●●

●

●●●●●●●●●●●●●●●●

●●●

●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●

●

●●●●●●●●●●●●●●●●●●●●●●●●

●

●●●

●●●●●●●●●●●●

●

●●●●●●●●●●●●●

●●

●●●●●●●●

●●●

●

●●●●●●●●●●●●

●●●

●●●●●●●●●●

●●●

●●●●●●●

●●●

●

●●

●

●●●●●●●●

●

●

●●●●●●●●●

●

●

●●●●●●●●●●●●

●

●

●●●●●

●

●

●●●●●●●●●●●●●●

●

●

●

●●

●●●●●●●●●●●●●●●●●

●●

●●

●●●●●

●

●●●●●●●

●●

●●

●

●

●

●

●

●

●●

●

●●●●

●●●

●●●●●

●●

●●●

●

●

●●●●●●●●●●●●●●

●

●●

●

●●●●●●●●

●

●●●●●●●●●●●●

●

●

●●

●

●

●●●●●●●●●●

●

●

●●●●

●

●

●

●●●●●●●●●●●●●●●●

●

●

●

●

●●●

●●●●

●●

●●●●●●●●●●●●●

●●

●●●●●●

●●●●

●

●

●●

●●●●●●

●●

●●●●●●●●●

●●

●●

●

●

●●●

●

●

●

●●●

●

●●

●

●●

●

●●●●●●●●

●

●

●

●●

●

●●●●●●●●

●●

●●●●●●●●●

●

●●●

●

●

●

●

●

●●

●●●●●

●

●●●

●

110

120

130

140

150

160

170

180

0 1 2 3 4VM #

Con

sum

ptio

n (W

att)

Node consumption statistics (Cluster edel, 128GB HDD)

110 W

175 W

servers are not energy efficient

how to reduce consumption with identical nodes

?

The principles

place VMs on the minimum number of nodes

turn off idle nodes

rince/repeat for the dynamic version

Best Fit Decrease (BFD)a simple heuristic to pack VMs

sort VMs in desc order for each VM pick the suitable node with the least remaining space

efficiency decreases when nb. of dimensions increase

2-pass dynamic VM packingto manage the running VMs

on saturated nodes, BFD to move away VMs on non-saturated nodes

on low-loaded nodes, BFD to move away VMs on heavily loaded nodes

1) address performance issues

2) address energy efficiency issues

many variations using threshold based systems, load predictions …

No packing 2-passes packing

0

25

50

75

100

0 500 1000 1500 2000 2500

% C

PU N

ode-

1Domain-1

0

25

50

75

100

0 500 1000 1500 2000 2500

% C

PU N

ode-

2

Domain-2

0

25

50

75

100

0 500 1000 1500 2000 2500

% C

PU N

ode-

3

Domain-3

0

25

50

75

100

0 500 1000 1500 2000 2500

% C

PU N

ode-

4

Time (in sec.)

Domain-4

0

25

50

75

100

0 500 1000 1500 2000 2500

% C

PU N

ode-

1

0

25

50

75

100

0 500 1000 1500 2000 2500

% C

PU N

ode-

2

0

25

50

75

100

0 500 1000 1500 2000 2500

% C

PU N

ode-

3

0

25

50

75

100

0 500 1000 1500 2000 2500

% C

PU N

ode-

4

Time (in sec.)boot time Domain-4 Domain-3 Domain-2 Domain-1

No packing 2-passes packing

0

25

50

75

100

0 500 1000 1500 2000 2500

% C

PU N

ode-

1Domain-1

0

25

50

75

100

0 500 1000 1500 2000 2500

% C

PU N

ode-

2

Domain-2

0

25

50

75

100

0 500 1000 1500 2000 2500

% C

PU N

ode-

3

Domain-3

0

25

50

75

100

0 500 1000 1500 2000 2500

% C

PU N

ode-

4

Time (in sec.)

Domain-4

0

25

50

75

100

0 500 1000 1500 2000 2500

% C

PU N

ode-

1

0

25

50

75

100

0 500 1000 1500 2000 2500

% C

PU N

ode-

2

0

25

50

75

100

0 500 1000 1500 2000 2500

% C

PU N

ode-

3

0

25

50

75

100

0 500 1000 1500 2000 2500

% C

PU N

ode-

4

Time (in sec.)boot time Domain-4 Domain-3 Domain-2 Domain-1

worthy ?

300

400

500

600

700

800

0 5 10 15 20 25 30 35 40

Ener

gy (W

atts

)

Time (min)Native execution IVS execution prototype

baseline IVS 2-pass VM packing

interesting gains, the node boot time was the bottleneck

simplistic model (identical nodes)

Consequences of having heterogeneous nodes ?

Consequences of having heterogeneous nodes ?

different performance different Power Usage Effectiveness

…

how to estimate the benefits of a migration ?

a specialised model

vector packing problem +performance model +power model +cost model +…

power modelsestimate the energy consumption

usually linear equations

model the static and the dynamic energy profile of the components

Wh(node) = α Hw +β

VM performance models

a model to map VM performance with their host

a neutral performance unit mips, ECU, GCU, …

perf(VM, N) = α %cpu +βn

multi-core CPUs, DDR3 memory, spinning HD, PUE / CUE, boot/shutdown time

VM template migration duration

migration payback time

hw. particularities

a fine-grain power model

workload particularities

coarse to fine grain optimisation

-16%

-47%

- 27%-7%

2Consume better2012 from a spatial to a temporal point of view

2013 2016

let existing and new data centres become energy adaptive

DC4Cities

0

25

50

75

100

16/01 17/01 18/01 19/01 20/01Date

rene

wabl

e po

wer (

%)

0

200

400

600

16/01 17/01 18/01 19/01 20/01Date

Powe

r (W

)

grid renewable part

sunroof PV production

renewable energies are intermittent

1000

1500

2000

2500

3000

23:00 05:00 11:00 17:00 23:00Time

Request/s

0.0

0.5

1.0

1.5

2.0

0 200 400 600 800Performance degradation (Req/s)

Pena

lty (e

uros

)

6 to 20 moonshot cartridges

penalty modelSLO

so are elastic and batch-oriented applications

align workload to renewable energies availability

DC4Cities

shape EASCs for sustainable profitability

forecasts

energyproviders

sustainable objectiveEasc

application descriptor

elected working modes

smart cityauthority

Easc...Carverweb service

videotranscoder

energy adaptiveapplications

min(penalty(SLO) + penalty(SMA) + price(E))pick WMs such as

Energy Adaptive Software Components

attached to an application

exhibit - working modes - SLO (cumulative or instant) - actuators

(see UCC’15 paper)

An automaton for each EASC model the behavior,

+ penalty functions for the SMA, the SLA

0

100

200

300

400

500

16/01 17/01 18/01 19/01 20/01Date

Watts

applicationE−learningG−indexingWebsite

0

200

400

16/01 17/01 18/01 19/01 20/01Date

Watts


0

200

400

16/01 17/01 18/01 19/01 20/01Date

Watts


baseline (satisfy perf )

“green” (max renewable)

carver

17/01/15 18/01/15 19/01/15 20/01/15

0

1000

2000

perf

green

Carver pe

rfgre

enCarv

er perf

green

Carver pe

rfgre

enCarv

er

Run

ning

cos

t (eu

ros)

expenseenergySLOSMA

Resulting running costs

RECAP114

no holy grail

115

master the problem

116

Resource management in IaaS cloud · 2017-01-30 · VM5 6 min(#onlineNodes) = 3 sol #1: 1m,1m,2m. dynamic schedulers quality at a price N3 cpu mem N1 cpu mem N4 cpu mem N2 cpu mem

Documents