Cloud Control with Distributed Rate Limiting

1
Cloud Control with Distributed Rate Limiting

• Barath Raghavan, Kashi Vishwanath
• Sriram Ramabhadran, Kenneth Yocum Alex
  C.Snoeren
• Offence
• Alex Kiaie Shiqi Chen

2
Problems and Flaws

• Global Random Drop
• Longer estimate interval VS overhead
• What happened to GRD at last?
• Central Token Bucket
• Flow Proportional Share
• Gossip Protocal
• Exceeding limit
• Evaluation on PlanetLab

3
Global Random Drop

• GRD definitely works better under shorter
  estimate interval, but how about the overhead
  generated by frequent communication?

4
Global Random Drop

• It seems this paper proposes GRD and drops it in
  the end without a good reason.
• Whats the purpose of proposing GRD then? Only to
  show FPS sucks in maintaining fairness?

5
Central Token Bucket

• Seems the authors design everything to
  approximate the performance of CTB. So why dont
  we just stick to CTB and keep life simple??

6
Flow Proportional Share
FTP works fine under Branch 5 or 7 for 500
limiters. But as the performance for Branch 1
worsens a lot when we have more than 400
limiters. What will the changing point be for
Branch 5 or 7? Or should we abandon the Gossip
Protocal?
7
Flow Proportional Share
We can see the FPS scheme is very stable in
500-ms estimate interval condition. BUT why half
of the time the aggregate rate is above the 10Mps
limit??
8
Flow Proportional Share

• Evaluation on PlanetLab?
• Weve already doubted the credibility of
  PlanetLab for millions of times in this class
• And the evaluation scale is really small

9
Flow Proportional Share
And we noticed some performance we did not see in
previous evaluation result of stable
implementations. How to explain this? Does it
imply that there could be more problem with real
network implementation?
10
Thank you!

• Any comment? Question?