Your SlideShare is downloading. ×

Art of Using Xen at Scale

26,897

Published on

0 Comments
6 Likes
Statistics
Notes
  • Be the first to comment

No Downloads
Views
Total Views
26,897
On Slideshare
0
From Embeds
0
Number of Embeds
12
Actions
Shares
0
Downloads
124
Comments
0
Likes
6
Embeds 0
No embeds

Report content
Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
No notes for slide

Transcript

  • 1. Art  of  Using  Xen  at  Scale   Pradeep  Vincent  Principal  Engineer,  Amazon  Web  Services      
  • 2. Agenda  •  Amazon  Web  Services  -­‐  Intro  •  Virtualiza@on  in  the  Cloud  •  Scale  Impact  on  Virtualiza@on  •  Xen  Enhancements..  •  Future  Features  –  Some  Thoughts  
  • 3. AMAZON  WEB  SERVICES  -­‐  INTRO    
  • 4. Amazon  Web  Services       “..delivers  a  highly  scalable  cloud  compu@ng   plaIorm  with  high  availability  and   dependability,  and  the  flexibility  to  enable   customers  to  build  a  wide  range  of   applica@ons.”  
  • 5. Amazon  Web  Services    •  Includes  more  than  20  Services   –  Compute  Services   •  EC2   –  Storage  Services   •  S3   •  EBS   –  Database  Services   –  Content  Delivery  Services   –  Messaging  Services   –  And  many  more..  •  More  Details  :  aws.amazon.com  
  • 6. EC2  Instance  Types   High   High   Cluster  Standard   Micro   Memory   Compute   Compute  Instances   Instances   Instances   Instances   Instances  
  • 7. VIRTUALIZATION  IN  THE  CLOUD  
  • 8. Virtualiza@on  and  Cloud  •  Virtualiza@on  a  key  enabler  for  Cloud   Compu@ng  •  Virtualiza@on  Provides..   –  Secure  Control  Plane     –  Fine  grained  resource  alloca@on   –  Cloud  Specific  Business  Logic  in  Dom-­‐0    
  • 9. Virtualiza@on  in  the  Cloud  •  Guest  OS(Kernel)  chosen  by  customer   –  Large  and  ever  changing  set  •  Aggressive  inges@on  of  Security  fixes   –  Security  is  paramount  •  Unique  Challenges  at  Scale..      
  • 10. SCALE  IMPACT  ON  VIRTUALIZATION  
  • 11. Scale  Ma_ers..  
  • 12.     Intra  Host  ‘Scaling’   Mul@-­‐tenancy    
  • 13. Mul@-­‐tenancy  •  Scaling  Challenges   –  Performance  Scaling:  I/O  performance  etc   –  Tools  Scaling:  XM  etc  •  Isola@on  Challenges   –  Performance  Isola@on   –  VM  Fault  Isola@on  
  • 14. Fault  Isola@on  
  • 15. VM-­‐level  Fault  Isola@on  •  Failure  related  to  one  VM  isolated  from  Peer   VMs  •  Good  Hypervisor  Level  Fault  Isola@on  in  Xen  •  End-­‐to-­‐End  Fault  Isola@on  Important   –  Including  Dom-­‐0  I/O  Stack   –  Including  Control  Plane  (e.g.  Xenwatch)  
  • 16. Performance  
  • 17. Consistency  of  Performance     Performance  Isola@on    
  • 18. Performance  Scaling  •  Cri@cal  for  Mul@-­‐tenancy  •  Par@cularly  I/O  Performance  •  Good  in  Xen  •  Deserves  con@nued  Focus..    
  • 19. I/O  Performance    Direct  AHached  Device  Performance  Good  But..     Loss  of  Control  and  Flexibility  Not  Great  
  • 20.     Para-­‐virtualized  I/O  Performance   Important    
  • 21. Reduce  Cost  
  • 22. Hardware  Enablement  •  Xen  Enablement  bit  ‘Slow’  for  some  hardware  •  Faster  ‘Produc@on’  quality  enablement  in  Xen   –  Help  Cloud  adopt  new  hardware  faster    
  • 23.    Upstream  Linux  IntegraOon  GREAT  !!  
  • 24. Interface  Compa@bility  
  • 25. Interface  Compa@bility  •  Applies  to  Many  Types  of  Interfaces   –  Paravirtualized  Driver  Interface     •  Blkfront  –  Blkback   •  NeIront  –  Netbk   –  Hypercall  Interface   –  Control  Plane  Interface  (XM  vs  XL)  •  Has  Been  Good  in  Xen     –  Deserves  con@nued  focus  
  • 26. XEN  ENHANCEMENTS..  
  • 27. Serve  Spiky  Workloads  For  Less  
  • 28. Micro  instance  in  EC2  
  • 29. Micro  Instance  and  CPU  Hogs  
  • 30. Large  ‘CPU  Time’  Bucket   Bigger  than  10s  of   millisecs   10s  of    milliseconds  
  • 31. Some  Thoughts  on  FUTURE  FEATURES    
  • 32. Mul@-­‐@er  Scheduling  ?  •  Lots  of  use  cases  in  the  Cloud  •  Classes  with  priori@za@on   –  Beyond  ‘CPU  Pools’   –  Similar  to  Linux  Scheduler  ?   –  Real  Time  Scheduling/Advanced  Func@onality  •  Preserve  exis@ng  Scheduler  Seman@cs  to  the   extent  possible  
  • 33. Dom-­‐0  Resource  Accoun@ng  •  CPU  consumed  in  Dom-­‐0  for  each  VM   –  Isn’t  accounted  against  the  VM  per  se   –  Fairness  Issues   –  Priori@za@on  Issues    
  • 34. Dom-­‐0  Resource  Accoun@ng  •  Possible  Solu@ons   –  Group  Scheduling  in  Dom-­‐0  –  Par@al  Solu@on   –  Xen  and  Dom-­‐0  Scheduler  Coopera@on  ??   •  Hypervisor  –  Dom-­‐0  Interface  Blurred..   –  Something  else  ??  •  Scheduler  deals  with  Threads..   –  One  Thread  serving  Mul@ple  VMs  not  great   –  ‘Thread  Aligned’  VM  Resource  Consump@on  Good  
  • 35. Key  Takeaways  •  Virtualiza@on  Key  For  Cloud  •  End-­‐to-­‐End  Fault  Isola@on  Important  •  Performance  Isola@on  And  Performance   Consistency  Important  •  Para-­‐virtualized  I/O  Performance  Important  •  Faster  ‘Produc@on’  Quality  Hardware   Enablement    •  Interface  Compa@bility  Cri@cal  
  • 36. SHAMELESS  MARKETING  
  • 37. We  are  Hiring  !  
  • 38. We  are  Hiring  !  We  are  Hiring  !  
  • 39. We  are  Hiring  !  We  are  Hiring  !   We  are  Hiring  !   aws.amazon.com/careers   We  are  Hiring  !  
  • 40. THANK  YOU  •  EMAIL:  PRADEEPV@AMAZON.COM  •  MORE  INFO:  WWW.AWS.AMAZON.COM  

×