© 2010 VMware Inc. All rights reserved

The Design and Evolution ofLive Storage Migration in VMware ESXAli Mashtizadeh, VMware, Inc.Emré Celebi, VMware, Inc.Tal Garfinkel, VMware, Inc.Min Cai, VMware, Inc.

2

Agenda What is live migration? Migration architectures Lessons

3

What is live migration (vMotion)?

Moves a VM between two physical hosts No noticeable interruption to the VM (ideally)

Use cases:• Hardware/software upgrades

• Distributed resource management

• Distributed power management

4

Virtual Machine

Live Migration

Virtual Machine

Disk is placed on a shared volume (100GBs-1TBs) CPU and Device State is copied (MBs)

Memory is copied (GBs)• Large and it changes often → Iteratively copy

Source Destination

5

Live Storage Migration

What is live storage migration?• Migration of a VM’s virtual disks

Why does this matter?• VMs can be very large

• Array maintenance means you may migrate all VMs in an array

• Migration time in hours

6

Live Migration and Storage Live Migration – a short history

ESX 2.0 (2003) – Live migration (vMotion)• Virtual disks must live on shared volumes

ESX 3.0 (2006) – Live storage migration lite (Upgrade vMotion)• Enabled upgrade of VMFS by migrating the disks

ESX 3.5 (2007) – Live storage migration (Storage vMotion)• Storage array upgrade and repair

• Manual storage load balancing

• Snapshot based

ESX 4.0 (2009) – Dirty block tracking (DBT) ESX 5.0 (2011) – IO Mirroring

7

Agenda What is live migration? Migration architectures Lessons

8

Goals

Migration Time• Minimize total end-to-end migration time

• Predictability of migration time

Guest Penalty• Minimize performance loss

• Minimize downtime

Atomicity• Avoid dependence on multiple volumes (for replication fault domains)

Guarantee Convergence• Ideally we want migrations to always complete successfully

9

Convergence

Migration time vs. downtime

More migration time → more guest performance impact More downtime → more service unavailability

Factors that effect convergence:• Block dirty rate

• Available storage network bandwidth

• Workload interactions

Challenges:• Many workloads interacting on storage array

• Unpredictable behavior

Migration Time Downtime

Dat

a R

emai

ning

Time

10

Migration Architectures

Snapshotting

Dirty Block Tracking (DBT)• Heat Optimization

IO Mirroring

11

Dest VolumeSrc Volume

ESX Host

Snapshot Architecture – ESX 3.5 U1

VMDKVMDKVMDK VMDK

12

Synthetic Workload

Synthetic Iometer workload (OLTP):• 30% Write/70% Read

• 100% Random

• 8KB IOs

• Outstanding IOs (OIOs) from 2 to 32

Migration Setup:• Migrated both the 6 GB System Disk and 32 GB Data Disk

Hardware:• Dell PowerEdge R710: Dual Xeon X5570 2.93 GHz

• Two EMC CX4-120 arrays connected via 8Gb Fibre Channel

13

Migration Time vs. Varying OIO

2 4 8 16 320

100

200

300

400

500

600

Snapshot

Snapshot

Workload OIO

Mig

ratio

n Ti

me

(s)

14

Downtime vs. Varying OIO

2 4 8 16 320

5

10

15

20

25

Snapshot

Snapshot

Workload OIO

Dow

ntim

e (s

)

15

Total Penalty vs. Varying OIO

2 4 8 16 320

50100150200250300350400450500

Snapshot

Snapshot

Workload OIO

Effe

ctiv

e Lo

st W

orkl

oad

Tim

e (s

)

16

Snapshot Architecture

Benefits• Simple implementation

• Built on existing and well tested infrastructure

Challenges• Suffers from snapshot performance issues

• Disk space: Up to 3x the VM size

• Not an atomic switch from source to destination• A problem when spanning replication fault domains

• Downtime

• Long migration times

17

Snapshot versus Dirty Block Tracking Intuition

Virtual disk level snapshots have overhead to maintain metadata Requires lots of disk space

We want to operate more like live migration• Iterative copy

• Block level copy rather than disk level – enables optimizations

We need a mechanism to track writes

18

Dest VolumeSrc VolumeVMDK VMDKVMDK

Dirty Block Tracking (DBT) Architecture – ESX 4.0/4.1

ESX Host

19

Data Mover (DM)

Kernel Service• Provides disk copy operations

• Avoids memory copy (DMAs only)

Operation (default configuration)• 16 Outstanding IOs

• 256 KB IOs

20

Migration Time vs. Varying OIO

2 4 8 16 320

100

200

300

400

500

600

700

800

SnapshotDBT

Workload OIO

Mig

ratio

n Ti

me

(s)

21

Downtime vs. Varying OIO

2 4 8 16 320

5

10

15

20

25

30

35

SnapshotDBT

Workload OIO

Dow

ntim

e (s

)

22

Total Penalty vs. Varying OIO

2 4 8 16 320

50

100

150

200

250

300

350

400

450

500

SnapshotDBT

Workload OIO

Effe

ctiv

e Lo

st W

orkl

oad

Tim

e (s

)

23

Dirty Block Tracking Architecture

Benefits• Well understood architecture based similar to live VM migration

• Eliminated performance issues associated with snapshots

• Enables block level optimizations

• Atomicity

Challenges• Migrations may not converge (and will not succeed with reasonable downtime)

• Destination slower than source

• Insufficient copy bandwidth

• Convergence logic difficult to tune

• Downtime

24

Block Write Frequency – Exchange Workload

25

Heat Optimization – Introduction

Defer copying data that is frequently written

Detects frequently written blocks• File system metadata

• Circular logs

• Application specific data

No significant benefit for:• Copy on write file systems (e.g. ZFS, HAMMER, WAFL)

• Workloads with limited locality (e.g. OLTP)

26

Heat Optimization – Design

Disk LBAs

27

DBT versus IO Mirroring Intuition

Live migration intuition – intercepting all memory writes is expensive • Trapping interferes with data fast path

• DBT traps only first write to a page

• Writes a batched to aggregate subsequent writes without trap

Intercepting all storage writes is cheap• IO stack processes all IOs already

IO Mirroring• Synchronously mirror all writes

• Single pass copy of the bulk of the disk

28

Dest VolumeSrc Volume

VMDKVMDK

IO Mirroring – ESX 5.0

ESX Host

VMDK

29

Migration Time vs. Varying OIO

2 4 8 16 320

100

200

300

400

500

600

700

800

SnapshotDBTMirror

Workload OIO

Mig

ratio

n Ti

me

(s)

30

Downtime vs. Varying OIO

2 4 8 16 320

5

10

15

20

25

30

35

SnapshotDBTMirror

Workload OIO

Dow

ntim

e (s

)

31

Total Penalty vs. Varying OIO

2 4 8 16 320

50

100

150

200

250

300

350

400

450

500

SnapshotDBTMirror

Workload OIO

Effe

ctiv

e Lo

st W

orkl

oad

Tim

e (s

)

32

IO Mirroring

Benefits• Minimal migration time

• Near-zero downtime

• Atomicity

Challenges• Complex code to guarantee atomicity of the migration

• Odd guest interactions require code for verification and debugging

33

Throttling the source

34

IO Mirroring to Slow Destination

0

500

1000

1500

2000

2500

3000

3500

Read IOPS Source

Write IOPS Source

Write IOPS Destination

Time (seconds)

IOPS

Start

End

35

Agenda What is live migration? Migration architectures Lessons

36

Recap

In the beginning live migration

Snapshot:• Usually has the worst downtime/penalty

• Whole disk level abstraction

• Snapshot overheads due to metadata

• No atomicity

DBT:• Manageable downtime (except when OIO > 16)

• Enabled block level optimizations

• Difficult to make convergence decisions

• No natural throttling

37

Recap – Cont.

Insight: storage is not memory• Interposing on all writes is practical and performant

IO Mirroring:• Near-zero downtime

• Best migration time consistency

• Minimal performance penalty

• No convergence logic necessary

• Natural throttling

38

Future Work

Leverage workload analysis to reduce mirroring overhead• Defer mirroring regions with potential sequential write IO patterns

• Defer hot blocks

• Read ahead for lazy mirroring

Apply mirroring to WAN migrations• New optimizations and hybrid architecture

39

Thank You!

40

Backup Slides

41

Exchange Migration with Heat

1 2 3 4 5 6 7 8 9 10

0

50

100

150

200

250

300

350

400

450

500

Hot Block

Cold Block

Baseline without Heat

Iteration

Data

Cop

ied

(MB)

42

Exchange Workload

Exchange 2010:• Workload generated by Exchange Load Generator

• 2000 User mailboxes

• Migrated only the 350 GB mailbox disk

Hardware:• Dell PowerEdge R910: 8-core Nehalem-EX

• EMC CX3-40

• Migrated between two 6 disk RAID-0 volumes

43

Exchange Results

Type Migration Time DowntimeDBT 2935.5 13.297

Incremental DBT 2638.9 7.557

IO Mirroring 1922.2 0.220

DBT (2x) Failed -

Incremental DBT (2x) Failed -

IO Mirroring (2x) 1824.3 0.186

44

IO Mirroring Lock Behavior

Moving the lock region1. Wait for non-mirrored inflight read IOs to complete. (queue all IOs)

2. Move the lock range

3. Release queued IOs

IILocked region:

IOs deferred until lock release IIIUnlocked region:

Write IOs to source only

IMirrored region:

Write IOs mirrored

45

Non-trivial Guest Interactions

Guest IO crossing disk locked regions

Guest buffer cache changing

Overlapped IOs

46

Lock Latency and Data Mover Time

0 111.071649896852 234.57560937731 403.7525319068390

200000

400000

600000

800000

DM

Cop

y (m

s)

0 111.071649896852 234.57560937731 403.7525319068390

20000

40000

60000

Lock

Lat

ency

(m

s)

Time (s)

No ContentionLock Contention

Lock Contention IO Mirroring Lock 2nd Disk

47

Source/Destination Valid Inconsistencies

Normal Guest Buffer Cache Behavior

This inconsistency is okay!• Source and destination are both valid crash consistent views of the disk

Time

Guest OSIssues IO

Source IOIssued

DestinationIO Issued

Guest OSModifies

Buffer Cache

48

Source/Destination Valid Inconsistencies

Overlapping IOs (Synthetic workloads only)

Seen in Iometer and other synthetic benchmarks File systems do not generate this

Virtual Disk

Source Disk Destination Disk

IO 1IO 2

IO 1IO 2 IO 1

IO 2Issue Order Issue Order

Issue Order

LBALBA

LBA

IO Reordering

49

Incremental DBT Optimization – ESX 4.1

Write to blocks Dirty block

? ? ? ? ?

Disk Blocks

Copy Ignore Copy Ignore