CloudAP: Improving the QoS of Mobile Applications with Efficient VM Migration

Renyu Yang, Ph.D. Student School of Computing, Beihang University

yangry@act.buaa.edu.cn In IEEE HPCC, Zhangjiajie, China, 2013

1

Joint work with Yunkai Zhang, Tianyu Wo, Chunming Hu, Junbin Kang and Lei Cui

Mobile Computing and Cloud • Mobile device

– the amount of smartphone users worldwide exceeded 1 billion in the third quarter of 2012 and continues to grow sharply.

– Shortcomings: low battery volume, poor resource configuration

• Cloud Computing – elastic resource management

– application hosting

– resource poverty compensation

2

• “Mobile + Cloud” – Virtualization techniques & Remote Execution

• Execution separates from display and the results are transmitted to the thin device client

• Cross-platform enhanced execution mechanism. Ex: make Windows application run on Android mobile phones

– Challenges: WAN(Wide Area Network) latency!

Motivation • How to mitigate the WAN latency?

– trusted, resource rich computers or clusters near mobile users -- Cloudlet proposed by Mahadev

– Initialize custom VMs on Cloudlet Software is run on cloudlets

– Drawbacks:

• the software can only be run on one single cloudlet

• discard the previous execution environment before initializing a new one.

3

• Requirements: – Pervasive and seamless continuous software service provision.

• Short service interruption time.

– service’s continuous execution

• Rapid and transparent execution environment switch.

• CloudAP

CloudAP

• Inspired by “Access Point” in communication areas.

• Small-scale cloud infrastructure close to the mobile devices and can be connected to Cloud datacenter

• CloudAP + Cloud Center

• Typical Scale

– Cloud Center – CloudAPs – Servers –VMs

– VM – User

• Example

4

Options CloudAP Cloud Center

Network LAN WAN

Sharing Few users Large numbers of users

Software Execution Yes Yes

Latency Low High

Bandwidth High Low

Scale Small Large

Scenario

5

Moving

How to make these functionalities come true ? How to maintain the execution environment by using an efficient mechanism?

Proposed Two-tier Architecture

6

VM Schedule Manager

AP Agent

Local scheduler

monitoring

Resource pool

Registry

VM Migration Controller

Cloud Agent

App VNC viewer VM VM

PM PM

VM VM VM VM

VMM VMM

Portal

Portal

Detect the geographical location Parse messages and send operation instructions

Design of Whole-system VM live migration

7

• Objective: shorten the switch time – The duration from when the migration initiates to when the VM runs

in the destination.

– Prompt execution environment switch into a better QoS surroundings.

• Solutions – COW(copy-on-write) file system: base + cow image

• Base image(large size) : read only and well-planned in advance.

• Cow image: the added or modified.

– Hybrid Copy Strategy • Disk : post copy but only for the cow image

• Memory : pre & post hybrid copy

– Considering the high frequency of access to memory pages, pure post-copy will lead to frequent memory page fault and performance degradation.

Design

8

Pre-copy memory single round

Suspend VM, migrate CPU statesSuspend VM, migrate CPU states

Resume the VM on destinationResume the VM on destination

Dirty memory post-copy

and whole disk post-copy

InitializationInitialization

Freeze-an

d-

copy

Post-co

py

Pre-co

py

Transfer the memory dirty-bitmap

• Only one iteration • Dirty-bitmap is used to record

the modified memory

• Pull on demand : dirty page pulling request occurs once a dirty page in the dirty-bitmap is accessed to.

• Push actively: the source pushes dirty memory pages to the destination according to the dirty-bitmap

Similarly, all the sectors in the disk image file should be set dirty in a block-bitmap

Optimization • Most dirty sectors are pushed in advance rather than pulled

when they are operated?

– Analyze the disk I/O characteristics

– Predict the I/O behaviors

– Push the suitable sectors intelligently

• Strong locality and continuity for read operations.

– If a read operation occurs to one dirty sector, the backward dirty sectors adjacent to this sector have a high probability to be accessed to.

• Data Compression Approach

– A long serial bit 0 or 1 in most dirty sectors.

– RLE9(Run-Length Encoding) strategy.

9

Experimental Setup

10

QEMU-KVM kvm-84 512MB memory 10GB base disk image 250MB cow image

16 dell servers with Intel(R) Core(TM) i7 860 2.80GHz CPU, 4GB memory and 320GB disk

VM

≤100Mbps

WiFi(54Mbps)

simulated 3G network (average 5Mbps bandwidth and 35 ms latency)

3 dell servers

Metrics

• Downtime *

– the time interval during which software service is entirely unavailable

• Switch time*

– the duration from when the migration is initiated to when the VM resumes and starts to run in the destination

• Total migration time

– the duration from when the migration starts to when the VM runs independently in the destination

• Amount of migrated data

11

QoS improvement in two-tier architecture

12

Operations Cloud center(s) CloudAP(s)

Page scrolling 1.08 0.54

Text input 0.32 <0.10

Text selection 0.51 <0.10

Color change 0.53 <0.10

Table insertion 0.52 <0.10

Text dragging 1.04 0.34

Text centering 0.32 <0.10

We emulate the typical behavior of a user who tries to use Abiword, a popular text editing remote executed software on his own mobile phone.

only takes less than 0.1s, which is acceptable for users

Efficient execution environment migration

13

Execution environment migration may cause a temporal downtime and a moment of turbulent service.

Workload Downtime(ms) Switch time(s)

Abiword 9.5 24.16

Openoffice 10 25.93

Firefox 9.6 29.91

• The service downtime can be limited in 10 milliseconds, which can be nearly neglected.

• The total VM switch time is limited within 30s, which is acceptable for user to sacrifice this time for obtaining a better software service

Migration performance comparison

14

TPM Deltas Ours TPM Deltas Ours TPM Deltas Ours

1

10

100

1000

10000

1955

9.5

2020

FirefoxAbiword

do

wn

time

(ms)

OpenOffice

downtime(ms)

21001930

10 9.6

30603020

TPM Deltas Ours TPM Deltas Ours TPM Deltas Ours

0

20

40

60

80

29.9

97.795.2

24.2

7473.7

25.9

75.675.1

switc

h tim

e(s)

switch time(s)

Abiword OpenOffice Firefox

Firstly realize the TPM[11](Three Phase Migration) and Deltas-apply[13] scheme in the hypervisor KVM

reduced significantly to only 10 milliseconds in comparison to 2-3s

Downtime Switch time

reaches an average 67% switch time reduction

15

0

10

20

30

40

50

60

70

80

90

100

110

34.81

74

FirefoxOpenOffice

time

cons

umpt

ion(

s)

Abiword

disk post synchronization

memory migration

disk precopy

73.976.06 75.6

39.12

102.93

97.7

43.15

TPM Deltas Ours TPM Deltas Ours TPM Deltas Ours

0

100

200

300

400

500

600

700

800

900

1000

1100

1200

437.91

817.04816.7

417.51

795.75793.7

OursDeltasTPM

FirefoxOpenOffice

amou

nt o

f mig

ratio

n da

ta(M

B)

disk synchronization

memory migration

disk precopy

Abiword

TPM Deltas Ours DeltasTPM Ours

1080.791107.34

481.66

Migration performance comparison

Migration time Migrated Data

Although takes a slight longer time(13.6s) during disk post synchronization, the total migration time is greatly reduced by at most 58% in all circumstance.

The entire migration data is only 481.66MB reaching roughly 55.4% and 56.5% decrease.

Conclusion

• Two-tier software execution architecture considering CloudAP and Cloud center (Public cloud).

• Optimized whole-system VM migration approach which fulfills rapid migration and effective execution environment switch.

16

Thank you for your attention!

Renyu Yang, Ph.D. student School of Computing Beihang University, Beijing, China Homepage: http://act.buaa.edu.cn/yangrenyu Email: yangry@act.buaa.edu.cn