Throughput Analysis of End-to-End Measurement Based Admission Control in IP

1
Throughput Analysis of End-to-End Measurement
Based Admission Control in IP

• G. Bianchi, A. Capone, C. Petrioli.
• Infocom 2000

2
Goals

• Achieve tight QoS control over the Internet
  without modifying its fundamental architectural
  principles.
• Whether an stateless, scalable architecture
  (DiffServ) can provide performance comparable to
  heavyweight per-flow resource management
  approaches.

3
EMBAC Operation

• Connection consists of two phases
• Probing Phase
• Data Phase
• Probing Phase
• Constant rate flow of packets with low priority,
  tagged Probing
• Measure packet arrival statistics over
  measurement interval Tm
• Based on measured statistics, indicate if
  resources available - send a Feedback packet
• If favorable, switch to Data phase.

4
(No Transcript)
5
EMBAC Features

• Core routers are stateless- only differentiate
  between classes of packets (Probing or Data)
• Probe packets do not compete with Data packets
  for bandwidth - router maintains two different
  queues.
• Implies that in case of high congestion Probe
  packets suffer considerably which is measures at
  the end nodes.
• Reduce Probe packet congestion by setting Probe
  packet life time (PLT)
• All accepted connections share same loss/delay
  performance
• Self stabilizing and robust
• EMBAC throughput lower than centralized decision
  based scheme

6
EMBAC Throughput analysis Constant Rate Sources

• Call arrivals, Poisson, rate ?/s
• Tm Constant duration of probing phase
• 1/?, average connection duration
• Each accepted connection has data rate B
• Bp Probing rate, Bt decision threshold rate
• Bp ? Bt ? B
• m(t) number of probing connections at time
  instant t
• K(t) number of accepted connections at time
  instant t

7
Analysis Contd.

• Instantaneous rate Br(t) , perceived at a probing
  connection end point is approximated by
• Use Approximation A1 Br ? Br(td)
• Then, due to acceptance condition Br ? Bt

8
Analysis Contd.

• We get the following constraint on m(t)
• Consider K(t), m(t) as a two-dimensional
  stochastic process
• Approx. 2 K(t) and m(t) are independent
• Number of probing connections has a poisson
  distribution
• and one can calculate Pa(K) probability of being
  admitted given K data connections
• And once can then model this system as a
  Birth-death Markov chain, and hence calculate the
  throughput.

9
Simulations
10
Simulations
11
Analysis VBR Notations

• B rate during ON period
• Nk(t) Number of active connections
• Ton Average period of ON state
• Toff Average period of OFF state
• x(Tm,t) average number of active sources found
  in time interval (t-Tm,t)

12
Simulations VBR
13
Simulations VBR
14
Simulations (VBR)
15
Simulations (VBR)
16
Simulations (VBR)