P1252428541JMSdj

1
A Measurement Based Admission Control Algorithm
for Integrated Services Packet Networks (
extended version)
S. Jamin, P. Danzig, S. Shenker, L. Zhang
2
Types of Services

• Guaranteed service
• a absolute bound on delay of every packet is
  provided
• Predictive service
• a fair but not absolute bound on delay of every
  packet is provided

3
Guaranteed service

• Provides hard or absolute bound on delay of every
  packet
• Sources are described by either peak or average
  rates or filter like token bucket
• Descriptions provide upper bounds on traffic
  generated by source
• Admission control algorithms use these a priori
  characterizations to calculate worst case
  behavior of flows
• Network utilization low when flows are bursty
• Network utilization acceptable when flows are
  smooth

4
Predictive Service

• Does not provide absolute bound on delay
• service commitment is somewhat less reliable
• sources characterized by token bucket filters at
  admission time
• behavior of existing flows determined by
  measurement rather than a priori
  characterizations
• when only few flows present, unpredictability of
  individual flows causes measurement based
  approach to be conservative
• delivers significant gain in utilization only
  when there is high degree of statistical
  multiplexing

5
Objective of paper

• Describes a measurement based admission control
  algorithm for predictive service
• To answer the following questions
• can reliable delay bounds be provided for
  measurement based admission control algorithm
• does predictive service increase network
  utilization

6
Measurement Based Admission Control

• Set of criteria controlling whether to admit a
  new flow
• based on approximate model of traffic flows
• uses measured quantities as inputs
• Measurement process to measure the inputs
• set of criteria used in this case
• delay bound as a result of admitting the new
  flow
• link utilization as a result of admitting the
  new flow

7
Measurement Based Admission Control

• Sources requesting services characterize worst
  case behavior
• to guarantee hard bound on delay
• worst case behavior considered
• conservative estimate of bound obtained
• new flow may not be admitted due to conservative
  estimate
• results in low utilization
• to achieve fairly reliable bound for predictive
  service
• approximate maximal delay by replacing worst
  case parameters with measured quantities
• results in higher utilization

8
Worst case delay

• Theorem by Parekh says the delay bound for a
  class j is the one time delay burst that accrues
  if the aggregate bucket of all classes 1 through
  j flows are simultaneously dumped into the switch
  and all classes 1 through j-1 sources continue to
  send at their reserved rates with j having lowest
  priority
• this worst case delay can change
• when flow of same class is admitted
• when flow of higher priority class is admitted
• when guaranteed flow is admitted
• classes classified according to the delay bounds
  tolerated by that class

9
Computation of new delay bound

• The delay bounds as a result of admitting the new
  flow is computed
• for e.g.,
• When new predictive flow a of same class is
  admitted
• new delay bound old delay bound k( bucket
  size of a)
• similarly new delay bounds are calculated for the
  case
• When new predictive flow a of lower class is
  admitted
• When new guaranteed flow a is admitted

10
Equivalent Token bucket filter

• In equations replace worst case values by
  measured values
• substitute the reserved rates by the measured
  rates
• substitute the worst case delays by the measured
  delays
• this results in describing the existing aggregate
  traffic by an equivalent token bucket filter with
  parameters determined from traffic measurement

11
Admission control algorithm

• New predictive flow if incoming flow a requests
  predictive service in a particular class
• request denied if sum of flows requested rate
  and current usage exceeds link utilization level
• request denied if admitting new flow violates
  delay bound of the same priority level
• request denied if admitting new flow violates
  delay bound of the lower priority level

12
Admission control algorithm

• New guaranteed flow if incoming flow a requests
  guaranteed service
• request denied if total bandwidth of the flows
  including new flow exceeds link utilization level
• request denied if admitting new guaranteed flow
  violates delay bound of predictive services
• this is due to the decrease in the bandwidth
  available for predictive flows due to admission
  of a guaranteed flow

13
Measurement mechanism Time window

• Two measurements
• experienced delay
• utilization
• to estimate delays measure the queuing delays d
  of every packet
• to estimate utilization sample the usage rate of
  the guaranteed service and the predictive class
  over sampling period of S packet transmission
  units

14
Measuring delay

• Measurement variable to estimate queuing delay is
  D
• Measurement window of T packet transmission units
  used as basic measurement block
• value of D updated on three occasions
• at end of measurement block D updated to reflect
  maximal packet delay seen in previous block
• when individual delay measurement exceeds
  estimated maximum queuing delay D is updated to
  times the sampled delay
• whenever a new flow is admitted D is updated to
  the value of projected delay from equation

15
Measuring rate

• Measurement variable to estimate queuing delay is
  V
• Measurement window of T packet transmission units
  used as basic measurement block
• value of V updated on three occasions
• at end of measurement block V updated to reflect
  maximal sampled utilization seen in previous
  block
• when individual utilization measurement exceeds
  estimated V, V is updated to the sampled value
• whenever a new flow is admitted V is updated

16
Performance Tuning

• Parameters , S, T can be varied for better
  performance
• smaller S, more sensitive to bursts , larger S
  smoother traffic
• smaller T means more adaptability to changes in
  traffic load, larger T results in more stability
• larger T results in fewer delay violations and
  lower link utilization

17
COPS

• A simple client/server model for supporting
  policy control over QoS signaling protocols and
  provisioned QoS resource management
• used to exchange policy information between a
  policy server (PDP) and its clients (PEP)
• for e.g., a policy client maybe RSVP routers that
  must exercise policy based admission control over
  RSVP usage

18
COPS

• A simple client/server model for supporting
  policy control over QoS signaling protocols and
  provisioned QoS resource management
• used to exchange policy information between a
  policy server (PDP) and its clients (PEP)
• for e.g., a policy client maybe RSVP routers that
  must exercise policy based admission control over
  RSVP usage