UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine...

38
XvMo%on: Unified Virtual Machine Migra%on over Long Distance Ali José Mash,zadeh, Min Cai, Gabriel TarasukLevin, Ricardo Koller, Tal Garfinkel, Sreekanth SeBy Stanford University – VMware, Inc.

Transcript of UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine...

Page 1: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

XvMo%on:  Unified  Virtual  Machine  

Migra%on  over  Long  Distance Ali  José  Mash,zadeh,  Min  Cai,  Gabriel  Tarasuk-­‐Levin,  

Ricardo  Koller,  Tal  Garfinkel,  Sreekanth  SeBy  Stanford  University  –  VMware,  Inc.  

Page 2: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Live  Migra%on

Virtual  Machine  

Source  Host   Des,na,on  Host  Cloud  

Page 3: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Migra%on  Benefits

• Test  to  produc,on  migra,ons  • Disaster  Preparedness  Tes,ng  • Cross-­‐Datacenter  Load  Balancing  • Shared-­‐Nothing  Architecture  

Page 4: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Virtual  Machine  

Migra%on  in  Prac%ce

Virtual  Machine  

Source  Host   Des,na,on  Host  

Requires:  Shared  Storage  Fast  Networking  Single  Subnet  

Page 5: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

This  is  not  what  we  really  want

• Migra,ons  are  limited  to  machines  that:  With  shared  storage,  fast  networks,  and  same  LAN  

•  Technological  Changes:  •  Shared  nothing  applica,on  architectures  •  Network  mobility  possible:  LISP,  OTV,  VXlan,  OpenFlow  (SDN)  

•  Very  large  virtualized  datacenters  

• No  reason  for  these  limita,ons  any  longer  • Customers  have  new  use  cases  

Page 6: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

XvMo%on

• XvMo,on:  First  commercially  viable  WAN  migra,on  

• Achieve  Low  Down,me  AND  Atomic  Switchover  • Uses  Asynchronous  IO  Mirroring  

• Principle:  Flow  control  mechanisms  for  memory  and  disk  

Page 7: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Customer  Scenario

LAN   WAN  Bandwidth   10  Gbps  

(some,mes  40  Gbps)  1  Gbps  or  less  

Latency   <1  ms   ~100  ms  Typical  Network   Dedicated  NIC(s)   Shared  connec,on  

between  sites  

Page 8: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Example  Workload

Applica,on   Highly  Available  Database  

Client  

HA  Timeouts  several  seconds  TCP  Timeouts  120  seconds  

Page 9: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Down%me  (Switchover  Time)

Virtual  Machine  

Running  Stopped  Running  

Goal:  Less  than  1  Second  of  Down,me  

Source  Host   Des,na,on  Host  

Page 10: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Virtual  Machine  

Atomicity

Virtual  Machine  

Source  Host   Des,na,on  Host  

Goal:  Atomic  switchover  

Page 11: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

XvMo%on

• Unifies  Memory  Migra,on  and  Storage  Migra,on  •  Tolerates  wide  area  network  bandwidth/latency  and  reliability  

•  Tolerates  heterogeneous  storage  performance  •  Down,mes  and  workload  impact  comparable  to  local  migra,on  

•  Atomic  switchover  

• Deployed  in  customer  metro  area  networks  • Cross  con,nent  migra,on  e.g.  Palo  Alto  to  India  is  prac,cal  

Page 12: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Overview

• Architecture  Overview  • Wide  Area  Memory  Migra,on  • Wide  Area  Storage  Migra,on  • Evalua,on  

Page 13: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Virtual  Machine  

Unified  Live  Migra%on

Virtual  Machine  

Source  Host   Des,na,on  Host  

Page 14: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

XvMo%on  Architecture

IO  Mirror  

Virtual  Machine  IO  

IO  Buffer  

Live  Migra,on  

Stream

s  

Stream

s  

Source   Des4na4on  

IO  Buffer/Writer  

Live  Migra,on  

TCP  

Page 15: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Overview

• Architecture  Overview  • Wide  Area  Memory  Migra4on  • Wide  Area  Storage  Migra,on  • Evalua,on  

Page 16: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

XvMo%on  Architecture:  Live  Migra%on

IO  Mirror  

Virtual  Machine  IO  

IO  Buffer  

Live  Migra,on  

Stream

s  

Stream

s  

Source   Des4na4on  

IO  Buffer/Writer  

Live  Migra,on  

TCP  

Page 17: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Live  Migra%on:  Total  Time  vs  Down%me

Memory  

Source   Des,na,on  

Page 18: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Live  Migra%on:  Total  Time  vs  Down%me

Memory  

Source   Des,na,on  

Page 19: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Live  Migra%on:  Total  Time  vs  Down%me

Memory  

Source   Des,na,on  

Page 20: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Live  Migra%on:  Total  Time  vs  Down%me

•  Itera,ve  copy  hopefully  reduces  the  working  set  each  itera,on  • Depends  on  Network  being  faster  than  Dirty  rate  

Memory  

Source   Des,na,on  

Page 21: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Current  Solu%on:  Early  Resume/Post-­‐copy • Problem:  Applica,ons  can  change  memory  faster  than  network  bandwidth  

•  Solu,on:  •  Stop  migra,on  copy  everything  •  Resume  early  if  down,me  is  high  

• Des,na,on  remote  page  faults  on  missing  pages  

Page 22: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Stun  During  Page  Send  (SDPS)

• Problems  with  Early  Resume:  •  Remote  page  faults  very  expensive  for  high  latency  networks  

•  Not  atomic:  Unsafe  for  WANs  where  network  hiccups  are  more  likely  

•  Solu,on:  ThroBle  VM  just  enough  to  keep  dirty  rate  <  network  rate  

Page 23: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

SDPS  Down%me  Results

0  

0.5  

1  

1.5  

2  

2.5  

24-­‐users   48-­‐users  

Down4

me  (s)  

SDPS  OFF   SDPS  ON  

Page 24: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Overview

• Architecture  Overview  • Wide  Area  Memory  Migra,on  • Wide  Area  Storage  Migra4on  • Evalua,on  

Page 25: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

XvMo%on  Architecture:  IO  Mirroring

IO  Mirror  

Virtual  Machine  IO  

IO  Buffer  

Live  Migra,on  

Stream

s  

Stream

s  

Source   Des4na4on  

IO  Buffer/Writer  

Live  Migra,on  

TCP  

Page 26: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Problem:  Synchronous  mirroring  over  the  WAN  slows  guest  workload   Goal:  Hide  network  latency  from  VM  

Page 27: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

VM  IO  

IO  

XvMo%on  Buffering:  Asynchronous  IO

TCP  Streams  Transport  Framework  

Source   Des4na4on  

Bulk  Disk  Copy  

Mirror  IO  

IO  Buffer  

Streams  Transport  Framework  IO  

Page 28: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

XvMo%on  Buffering:  Asynchronous  IO

VM  TCP  

Streams  Transport  Framework  

Source   Des4na4on  

IO  Buffer  

Streams  Transport  Framework  

IO   IO   IO   IO  

Page 29: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Problem:  Workload  may  be  too  fast  on  source  for  the  des;na;on   Goal:  ThroBle  copy  process  and  workload  as  necessary  

Page 30: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

IO  

XvMo%on  Buffering:  Conges%on  Control

VM  TCP  

Streams  Transport  Framework  

Source   Des4na4on  

IO  Buffer  

Streams  Transport  Framework  

IO  

IO   IO   IO   IO   IO   IO   IO   IO  Back  Pressure  

Limit  per-­‐disk  OIOs/buffer  on  des,na,on  

Send  free  buffer  space  

Slow  Disk  

Page 31: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

IO  

XvMo%on  Buffering:  Conges%on  Control

VM  TCP  

Streams  Transport  Framework  

Source   Des4na4on  

IO  Buffer  

Streams  Transport  Framework  

IO  

IO   IO   IO   IO   IO   IO   IO   IO  Back  Pressure  

Limit  per-­‐disk  OIOs/buffer  on  des,na,on  

Send  free  buffer  space  

Slow  Disk  

Page 32: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Overview

• Architecture  Overview  • Wide  Area  Memory  Migra,on  • Wide  Area  Storage  Migra,on  • Evalua4on  

Page 33: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Evalua%on

• Migra,on  Time:  Total  end-­‐to-­‐end  ,me  • Down,me:  Time  machine  execu,on  is  suspended  for  final  switchover  • Workload  Penalty:  Average  percentage  penalty  to  the  VM  workload  

• OLTP  IO  Workload  (data  disk  only)  •  10  GB  system/12  GB  data  

Page 34: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

XvMo%on  Down%mes

0  

0.2  

0.4  

0.6  

0.8  

1  

1.2  

0   50   100   150   200  

Down,

me  (s)  

Latency  (ms)  

Take  Away:  ~1  second  down,me  independent  of  distance  

Page 35: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

XvMo%on  Migra%on  Time

0  

50  

100  

150  

200  

250  

300  

0   50   100   150   200  

Total  Tim

e  (s)  

Latency  (ms)  

Latency  

Take  Away:  small  linear  ,me  increase  with  distance  

Page 36: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

California  to  India  Migra%on

•  1  Gbps  network  with  214  ms  RTT  

• OLTP:  68  MB/s  disk  copy  –  89  MB/s  memory  copy  •  ~11%  Workload  Penalty  from  ThroBling  

Page 37: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Summary

•  XvMo,on  frees  migra,on  from  the  need  for  shared  storage  and  fast  local  networks  •  Tolerates  wide  area  network  bandwidth/latency  and  reliability  

•  Tolerates  heterogeneous  storage  performance  •  Down,mes  and  workload  penalty  comparable  to  local  migra,on  

•  Atomic  Switchover  

•  Enables  new  use  cases  –  e.g.  disaster  preparedness,  cluster  upgrade,  shared  nothing  

•  On  the  path  to  deployment:  •  Deployed  in  customer  metro  area  networks  •  Cross  con,nent  migra,on  e.g.  Palo  Alto  to  India  is  prac,cal  

Page 38: UnifiedVirtualMachine MigraonoverLongDistance · XvMo%on: UnifiedVirtualMachine MigraonoverLongDistance Ali$José$Mash,zadeh,$Min$Cai,$Gabriel$TarasukLevin,$ Ricardo$Koller,$Tal$Garfinkel,

Ques%ons?