energy-efficient virtual machine provision algorithms for cloud system ching-chi lin institute of...

23
Energy-efficient Virtual Machine Provision Algorithms for Cloud System Ching-Chi Lin Institute of Information Science, Academia Sinica Department of Computer Science and Information Engineering, Nation Taiwan University Pangfeng Liu Department of Computer Science and Information Engineering, Nation Taiwan University Graduate Institute of Networking and Multimedia, Nation Taiwan University Jan-Jan Wu Institute of Information Science, Academia Sinica

Upload: erica-holmes

Post on 16-Dec-2015

213 views

Category:

Documents


0 download

TRANSCRIPT

Energy-efficient Virtual Machine Provision Algorithms for Cloud SystemChing-Chi LinInstitute of Information Science, Academia Sinica

Department of Computer Science and Information Engineering, Nation Taiwan University

Pangfeng LiuDepartment of Computer Science and Information Engineering, Nation Taiwan University

Graduate Institute of Networking and Multimedia, Nation Taiwan University

Jan-Jan WuInstitute of Information Science, Academia Sinica

IntroductionDatacenters consume

tremendous amount of energy.To reduce the power consumption

in a datacenter◦Virtual machine consolidation.◦Reduce the number of power-on idle

servers.

VM consolidationDefinition:

◦Virtual machines with different hardware requirements must be deployed using the minimum number of physical servers.

VM consolidation is an NP-hard problem.◦Can be reduced by bin-packing

problem.

Heuristic Solutions for VM ConsolidationFirst-Fit

◦Deploy VMs to the fist server that can accommodates this VM.

◦Load imbalance for power-on servers.

Round-Robin◦Deploy VM to the servers in circular

order. Balancing the load between servers.

◦Hard for physical servers to become idle.

Dynamic Round-RobinMinimize the number of physical

server used to run all virtual machines.

Extension from Round-RobinAdd two rules to reduce the

number of physical server in use.

Rule 1Physical server become “retiring”, i.e.,

not accepting new VMs, after one of its hosting VM finishes.

A B C

D E

G

F

H H

Rule 2If a retiring server stay active for

too long, migrates its hosting VMs and power off this server.◦Retirement threshold: the waiting

time before migrating hosting VMs.

A C

D E

G

F

A Hybrid MethodHybrid method chooses from

these two strategies depending on incoming traffic.◦During rush hours, Dynamic Round-

Robin may not perform better than First-Fit. Rush hour: a period of time with high VM

incoming rate.

◦Rush hour, use First-Fit.◦Otherwise, use Dynamic Round-

Robin.

System ModelsPower consumption model

◦The power consumption of a physical server is a linear function of its CPU load.[1][2]

◦The idle power is about 50% of the peak power.[1][2]

Migration model◦Live migration

[1]:Xiaobo Fan, Wolf-Dietrich Weber, and Luiz Andre Barroso. Power provisioning for a warehouse-sized computer.[2]:Gong Chen, Wenbo He, Jie Liu, Suman Nath, Leonidas Rigas, Lin Xiao, and Feng Zhao. Energy-aware server provisioning and load dispatching for connection-intensive internet services.

Verification

Number of threads(cores)

Power ModelThe assumption of linear power model

with 50% idle power is reasonable.The power consumption P of a physical

server:

α: the percentage of idle power versus the peak power.

c: the total number of cores required. c ≦ C C: the total number of cores on a physical machine. c/C: the load of a physical machine. Pp: peak power.

pPC

cP

1

Live MigrationMove a running VM to a different

physical server without interrupting the VM.◦Does not lengthen the execution

time of a VM.◦But slightly increase power

consumption of source and destination server. An extra load caused by migration

increases about 20% load to the destination server.

Migration ModelEnergy consumption on migrating VM:

tM: time migration takes. L: extra load caused by the migration process. Es, Ed: energy consumption on source and

destination physical machine. c/C, c’/C: the load of physical machines. c, c’ ≦ C α: the percentage of idle power versus the peak

power. Pp: peak power.

MpMp

dsm

tPLC

ctP

C

c

EEE

1'

1

Experimental SettingPhysical server:

◦Intel(R) Xeon(R) E5620 CPUs 2.40GHz, each CPU has four cores.

◦24GB memory.Virtual machine:

◦Small: 1v-core, 1.7 GB of memory.◦Large: 2 v-cores, 7.5 GB of memory.◦Extra large: 4 v-cores, 15 GB of

memory.

Experimental Setting(Cont.)Test cases

◦2,000 virtual machines.◦Arrival time follows normal

distribution, ranging from 0:00 to 23:59, to simulate the workload of a day. Rush hour: 10:00~14:00

◦Execution time is generated from uniform distribution, ranging from 2 to 12 hours.

Experiment 1 – Different Retirement ThresholdRetirement threshold

◦Time as threshold:0, 5, 10, 30 minutes

◦Physical server load as threshold: ≦50%

Baseline: First-Fit

Result 1 Using physical server load as

threshold consume the least power.

Experiment 2 – Performance ComparisonDynamic Round-Robin

◦Use physical server load as retirement threshold.

Baseline: First-Fit

Result 2The Hybrid method consume

about 3% less average power than First-Fit.

Experiment 3 – Comparison with EucalyptusEucalyptus strategies

◦Greedy & Round-Robin Does not power off idle physical servers.

◦Power-Save Power off idle physical servers.

Baseline: Eucalyptus Round-Robin

Result 3Dynamic Round-Robin and Hybrid

consume about 40% lesser compare to Eucalyptus default strategy.

ConclusionWe propose Dynamic Round-

Robin and Hybrid, to deploy the virtual machines for power saving purpose.

Both experiment and simulation results show that Dynamic Round-Robin and Hybrid we propose can reduce the average power consumption compared to other deployment algorithms with power-saving.

Thank you