energy efficient resource management in virtualised datacenters

Post on 02-Jul-2015

307 views 0 download

description

A coarse grain overview of energy-aware resource management approaches since the last ten years through 2 EU founded projects. Presented during the 2014 "Energy Aware Network" Labex day at Inria. (see http://www.ucnlab.eu/fr/node/66)

Transcript of energy efficient resource management in virtualised datacenters

energy efficient resource management in virtualised datacenters

Fabien Hermenier

(Environmental Protection Agency, 2007)

of the 2005 budget1.5%

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)

the brown constant

110 W

175 W

2003“let’s turn off useless stuff”elnozahy et al. :

Packingconcentrate WWW

requests

2006 “look! look it’s moving”Hansen et al. :

live migrationN1 N2VM1

1) model power consumption 2) pack VMs 3) turn-off unused nodes 4) minimize(W) 5) profit

consume less, the theory

hw. heterogeneity env. capabilities performance vs. energy workload volatility data center size

,,consume less, the practice

BtrPlace proposalcore reconfiguration algorithm

users scripts

+ =specialized reconfiguration algorithm

spread(VM[2..3]); preserve({VM1},’ucpu’, 3); offline(@N4);

0’00  to  0’02:  relocate(VM2,N2)  0’00  to  0’04:  relocate(VM6,N2)  0’02  to  0’05:  relocate(VM4,N1)    0’04  to  0’08:  shutdown(N4)  0’05  to  0’06:  allocate(VM1,‘cpu’,3)  

The reconfiguration plan

BtrPlace

the core reconfiguration algorithm is modeled wrt. the impact of actions on resources

allDifferent(dhost

1 , d

host

2 ) ^d

host

1 = c

host

2 ! d

st

1 � c

ed

2 ^d

host

2 = c

host

1 ! d

st

2 � c

ed

1

spread({VM1,VM2}):

Premade variables, constraints

Constraint Programming !

!

!

to the rescue

2010 2013

Energy aware ICT optimization policies (+ btrPlace)

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

energy-related variables are linked to core ones

energy-oriented constraintsMaxServerPower

DelayBtwMigrationsDelayBtwServerSwitch

PayBackTime

SpareCPUs

minEnergyCons

minGasEmissions

cap consumption

reduce ping-pong effects

migration as an investment

control the consolidation aggressiveness

optimisation criteria

500 1000 1500 2000 2500

1

234

5

6

7

Servers

Tim

e (m

inut

es)

P4G P4G + spare P4G + spare + vcpu P4G + spare + vcpu + delay

the core problem dominates scalability

coarse to fine grain optimisation

-16%

-47%

- 27%-7%

Consuming lessdealing with

THE BROWN CONSTANT

Consuming lessdealing with

UNAPPROPRIATE

HARDWARE

hw. community did not

chill

energy proportional servers, free cooling,

fanless processors

?workload agnostic

need to revamp software approaches migration

a balancing problem

2Consume better2012

2013 2016

let existing and new data centres become energy adaptive

DC4Cities

align workload to renewable energies availability

DC4Cities gray box approach

forecasts1) 24h power budget

2) alt. working modes

3) selected mode

energy adaptive applications

Energy adaptive Batch scheduling

How many parallel jobs, which ones to defer

Trade replicas against latency

Allocate resources against SLAs

Energy adaptive WWW

Energy adaptive IaaS

Energy adaptive …

ConclusionS

consume less, consume better, non-renewable power sources, DVFS, live-migration, deferrable workload, so many facets, non-deferrable workload, elasticity, VM packing, VOVO, VM balancing, white-box approach, black-box approach, solutions need to follow new capabilities and usage, priority over SLA, priority over savings, renewable power sources, fine grain power model, coarse grain power model, steady workload, bursty workload

in 20101.1 - 1.5 %

(Environmental Protection Agency, 2013)

USA

.org