Optimizing Live Migration of Virtual Machines across Wide Area Networks using Integrated Replication and Scheduling Sumit Kumar Bose, Unisys Scott Brock,

Optimizing Live Migration of Virtual Machines across Wide Area Networks using Integrated Replication and Scheduling

Sumit Kumar Bose, UnisysScott Brock, UnisysRonald Leaton Skeoch, UnisysNisaruddin Shaikh, UnisysShrisha Rao, IIIT-Bangalore

Agenda

Motivation and Problem Addressed

Proposed Solution

Experimental Results

Conclusion

Motivation and Problem Addressed

Migration Latency

VM Size

Threshold

Proposed Solution – CloudSpider

Proposed Solution

Hiber-waking vs

MigrationMigration Latency

VM Size

Threshold

Contd…

• Integrated replication and scheduling of VMs.– Selectively choose appropriate clouds for VM replication– Using past cost history

• However, blindly replicating VM images leads to increased storage requirements.

• Reduce overall storage requirements due to VM image replication by exploiting commonality across different VM images using de-duplication.– Past studies show that different VM images have significant

common content due to common substrates (OS, web-servers etc).

Contd…

• VM1 and VM2 images share significant similarity.

VM1

VM1

VM1

VM2

VM2VM2

Storage cost shoots up!!!!

Contd…

• Exploit de-duplication to reduce storage cost

VM1

VM1

VM1VM2

VM2

VM2

Trade-off between storage cost and operational cost

Contd…

• The research questions that arise in such contexts are to determine the – List of eligible sites for different VM images (Replication

phase) – Placement of the replicas at different cloud sites (Replication

phase)– Primary copy from among the replicas of a VM image

(Scheduling phase)

• Our primary focus is to develop efficient replication strategies for questions (i) and (ii) using our integrated VM replication and scheduling architecture

Algorithm

Line -1: for each site j in J do: Line -2: for each VM image i in Ij do: Line -3: Calculate

j i j

j k ki I k K k K

Sav size size

Line -4: Calculate ' '( ' & ' ) ( ) ( & ' )

ij

j i j i

r k ki I i i k K k K K i i

Sav size size

Line -5: Calculate ijj rSav Sav

Line -6: endfor Line -7: endfor Line -8: Let arg min ( )

ij ijij r j rr Sav Sav and assume ijr to be member

of the set Ri” Line -9: If max

"| |i iR N remove replica rij and update jSav and ijrSav ,

i in Ij and update Ij. Repeat the step at Line-8, until all the number of replicas for all the VM images are within the bound max

iN .

Experiments

• We implemented the replication strategy using Java.

• We extended the open-source CloudSim toolkit and incorporated the replication module in it.

• We experimented extensively and simulated different scenarios using the CloudSim toolkit

• To evaluate the performance of the CloudSpider algorithm we implemented the following naïve strategies:

– RandomMax – RandomMin – MaxReplica– MinReplica.

Contd…

– RandomMax and RandomMin strategies randomly select sites from the eligible list of a VM image for placing the replicas.

– MaxReplica and MinReplica strategies select the first few sites from the eligible list of VM for creating the replicas.

• Performance criteria: The average (over 100 simulation iterations) of total storage consumed by all the replicas belonging to all the VM images.

Contd…

• Subsequent tables show the variation in the average storage requirement by the different strategies when the following parameters is varied:– number of data-centers – number of VM images – file size – latency threshold– savings threshold.

Contd…

• Storage requirements with 10 VM images and image size is 1 MB and latency threshold = 0.99 & savings threshold = 0.25

StrategyNumber of DC

Number of VM Images

File Size (MBs)

Minimum number of Replicas

Maximum number of Replicas

Latency threshold

Savings threshold

Average Time (ms)

Average number of

replicas

Average storage requirement for all apps incl replicas

before dedup (MB)

Average storage requirement for all apps incl replicas after dedup (MB)

CloudSpider 5 10 1 1 2 0.99 0.25 526 2 20 4.31RandomMax 5 10 1 1 2 0.99 0.25 174 2 20 9.70MaxReplica 5 10 1 1 2 0.99 0.25 163 2 20 4.34RandomMin 5 10 1 1 2 0.99 0.25 104 1 10 7.06MinReplica 5 10 1 1 2 0.99 0.25 102 1 10 2.17

Contd…

• Storage requirements with 20 VM images and image size is 1 MB and latency threshold = 0.99 & savings threshold = 0.25

StrategyNumber of DC

Number of VM Images

File Size (MBs)

Minimum number of Replicas

Maximum number of Replicas

Latency threshold

Savings threshold

Average Time (ms)

Average number of

replicas

Average storage requirement for all apps incl replicas

before dedup (MB)

Average storage requirement for all apps incl replicas after dedup (MB)

CloudSpider 5 20 1 1 2 0.99 0.25 1407 2 40 3.75RandomMax 5 20 1 1 2 0.99 0.25 420 2 40 10.58MaxReplica 5 20 1 1 2 0.99 0.25 461 2 40 4.56RandomMin 5 20 1 1 2 0.99 0.25 291 1 20 9.50MinReplica 5 20 1 1 2 0.99 0.25 319 1 20 2.39

Contd…

• Storage requirements for RandomMax when different parameters to the problem are varied

Number of DC

Number of VM Images

File Size (MBs)

Minimum number of Replicas

Maximum number of Replicas

Latency threshold

Savings threshold

Average Time (ms)

Average number of

replicas

Average storage requirement for all apps incl replicas

before dedup (MB)

Average storage requirement for all apps incl replicas after dedup (MB)

5 20 2 1 2 0.99 0.25 1043 2 80 21.385 10 2 1 2 0.99 0.25 447 2 40 19.575 10 1 1 3 0.99 0.25 244 3 30 10.51

5 20 1 1 3 0.99 0.25 604 3 60 10.685 10 1 2 3 0.99 0.25 239 3 30 10.55

5 10 1 1 2 0.90 0.25 201 1.999 19.99 9.845 10 1 1 2 0.75 0.25 183 1.987 19.87 9.77

5 10 1 1 2 0.50 0.25 152 1.815 18.15 9.555 10 1 1 2 0.90 0.50 163 2 20 9.67

5 10 1 1 2 0.90 0.75 163 1.999 19.99 9.87

Contd…

• Storage requirements for RandomMin when different parameters to the problem are varied

Number of DC

Number of VM Images

File Size (MBs)

Minimum number of Replicas

Maximum number of Replicas

Latency threshold

Savings threshold

Average Time (ms)

Average number of

replicas

Average storage requirement for all apps incl replicas

before dedup (MB)

Average storage requirement for all apps incl replicas after dedup (MB)

5 20 2 1 2 0.99 0.25 543 1 40 19.06

5 10 2 1 2 0.99 0.25 233 1 20 13.955 10 1 1 3 0.99 0.25 105 1 10 7.14

5 20 1 1 3 0.99 0.25 270 1 20 9.725 10 1 2 3 0.99 0.25 103 1 10 6.84

5 10 1 1 2 0.90 0.25 104 1 10 7.015 10 1 1 2 0.75 0.25 99 0.999 9.99 7.035 10 1 1 2 0.50 0.25 99 0.963 9.63 6.85

5 10 1 1 2 0.90 0.50 99 1 10 7.065 10 1 1 2 0.90 0.75 101 1 10 7.05

Contd…

• Storage requirements for MaxReplica when different parameters to the problem are varied

Number of DC

Number of VM Images

File Size (MBs)

Minimum number of Replicas

Maximum number of Replicas

Latency threshold

Savings threshold

Average Time (ms)

Average number of

replicas

Average storage requirement for all apps incl replicas

before dedup (MB)

Average storage requirement for all apps incl replicas after dedup (MB)

5 20 2 1 2 0.99 0.25 1004 2 80.00 9.375 10 2 1 2 0.99 0.25 559 2 40.00 8.79

5 10 1 1 3 0.99 0.25 216 3 30.00 6.535 20 1 1 3 0.99 0.25 613 3 60.00 6.935 10 1 2 3 0.99 0.25 216 3 30.00 6.69

5 10 1 1 2 0.90 0.25 156 1.998 19.98 5.805 10 1 1 2 0.75 0.25 160 1.981 19.81 7.52

5 10 1 1 2 0.50 0.25 152 1.837 18.37 8.865 10 1 1 2 0.90 0.50 157 2 20 5.78

5 10 1 1 2 0.90 0.75 157 1.999 19.99 5.83

Contd…

• Storage requirements for MinReplica when different parameters to the problem are varied

Number of DC

Number of VM Images

File Size (MBs)

Minimum number of Replicas

Maximum number of Replicas

Latency threshold

Savings threshold

Average Time (ms)

Average number of

replicas

Average storage requirement for all apps incl replicas

before dedup (MB)

Average storage requirement for all apps incl replicas after dedup (MB)

5 20 2 1 2 0.99 0.25 613 1 40 4.715 10 2 1 2 0.99 0.25 248 1 20 4.40

5 10 1 1 3 0.99 0.25 110 1 10 2.285 20 1 1 3 0.99 0.25 320 1 20 2.255 10 1 2 3 0.99 0.25 111 1 10 2.19

5 10 1 1 2 0.90 0.25 115 1 10 3.075 10 1 1 2 0.75 0.25 122 1 10 4.12

5 10 1 1 2 0.50 0.25 112 1 10 5.625 10 1 1 2 0.90 0.50 109 1 10 3.10

5 10 1 1 2 0.90 0.75 114 1 10 2.98

Contd…

Number of DC

Number of VM Images

File Size (MBs)

Minimum number of Replicas

Maximum number of Replicas

Latency threshold

Savings threshold

Average Time (ms)

Average number of

replicas

Average storage requirement for all apps incl replicas

before dedup (MB)

Average storage requirement for all apps incl replicas after dedup (MB)

5 20 2 1 2 0.99 0.25 3479 2 80 7.59

5 10 2 1 2 0.99 0.25 1546 2 40 8.265 10 1 1 3 0.99 0.25 588 3 30.00 5.97

5 20 1 1 3 0.99 0.25 1571 3 60 5.395 10 1 2 3 0.99 0.25 617 3 30.00 6.245 10 1 1 2 0.90 0.25 488 2 20 4.42

5 10 1 1 2 0.75 0.25 435 1.975 20 5.865 10 1 1 2 0.50 0.25 400 1.9 20 7.54

5 10 1 1 2 0.90 0.50 506 1.999 19.99 4.375 10 1 1 2 0.90 0.75 480 1.996 19.97 4.50

• Storage requirements for CloudSpider when different parameters to the problem are varied

• Results: Cloudspider shows significant reduction in cost of storage requirement

Conclusion

• To overcome migration latencies associated with moving large sized files over the low-bandwidth networks, we propose to combine VM replication with VM scheduling.

• We compensate the additional storage requirement due to the increase in the number of replicas by exploring commonalities across different VM images using de-duplication techniques.

• The proposed architecture for integrated replication and scheduling – CloudSpider – is promising.

Page 22

Q & A

QU E S T I O N S?