Scaling Network Emulation Using Topology Replication

1
Scaling Network Emulation Using Topology
Replication

• Second Year Project
• Advisor Amin Vahdat
• Committee Jeff Chase, Jun Yang

2
Outline

• ModelNet
• Scalability
• Replication
• Motivation
• Theory
• Evalution
• Conclusions and Future Work

3
ModelNet

• Model network links as pipes
• Pipe is a queue
• Bandwidth, loss rate, latency,queuing discipline
  etc.
• Multiple flows share the same pipe(capture
  congestion)

4
Replication

• Emulation capacity limited by the core
• For a fixed topology measured in terms of packets
  per second processed
• Need more processors to scale
• Key observation Not one flow is responsible for
  the breakdown but a lot of them collectively
• Intuition Somehow let different processors
  handle different flows.
• The challenge How to synchronise so that the
  processor in unison emulate the behaviour expected

5
System Design
6
State diagram for a pipe

• Backoff / Set timer, bwbw b

FREE
CONSTRAINED
Timeout / Reset bw bw_max
State diagram for a pipe
7
Evaluation

• Long Lived Flows
• Short Web like flows
• What to measure ?
• Correctness/Accuracy
• Overhead

8
Web like flows
Client
Client
9
How to measure correctness

• Measure file access times for 10 files each from
  10 clients
• Network measurementgt large variance in resultsgt
  use Statistics
• Perform the test a large number of times(gt30)gt
  100 distributions
• Repeat with replication ongt100 distributions
• Compare pairwisegt 100 comparisons
• What does comparison mean?

10
Comparing Distributions

• Kolmogorov-Smirnov Test (KStest)
• Null Hypothesisgt CDFaCDFb
• Set significance level 0.05
• Unable to reject null hypothesis-gt with large
  confidence CDFaCDFb
• Results gt unable to detect for 90 pairs(90)

11
Comparing Distributions

• For at least 90 the cdfs seem to match.(Null
  hypothesis not rejected)
• For remaining 10
• Compare 90-ile of bandwidths seen
• Calculating percentage deviation of cfm compared
  to logical pipe
• Plot/quote error in throughput for these cases
• Other ways?

12
Q-Q plot

• A straight line here means that both samples are
  drawn from the same underlying distribution

13
Percentage error for 90-iles

• The cdf is good if the 90-ile has a low
  percentage error

14
Scalability

• Havent gone into this at all. Scaling with more
  number of cores?
• possibly graph showing more pps when the
  partitioning case has a lot of cross-traffic

15
Overhead

• Have a loose bound on the communication overhead
  which does not seem bad but more can be future
  work

16
Conclusions and Future Work

• Showed replication as an effective technique to
  scale.
• Parameters are application specific but more work
  required to quanitfy their roles.
• Partitioning along with replication should be
  considered
• A more realistic application evaluated

17
Thanks!

• Questions?