Building a Controlled Delay Assured Forwarding Class in DiffServ Networks

1
Building a Controlled Delay Assured Forwarding
Class in DiffServ Networks

• Parag Kulkarni
• Nazeeruddin Mohammad
• Sally McClean
• Gerard Parr
• Michaela Black
• Bryan Scotney
• School of Computing and Information Engineering
• Faculty of Engineering
• University of Ulster Coleraine, Northern Ireland
• parag_at_infc.ulst.ac.uk

2
Overview

• Quality of Service (QoS) within a typical router
  in the internet is quantified through several
  parameters
• e.g. Delay, Delay Variation, Packet Loss
• Effective Queue Management helps to improve QoS
• Objective of this research
• To design a ProActive Queue Management strategy
  that provides quantitative controlled delay
  guarantees on a per hop basis across an
  end-to-end (e2e) path by regulating queue size
  around an operating point q0 (thereby regulating
  delay below the control target and minimizing
  delay variation).

3
Background

• Service differentiation essential to support SLAs
• DiffServ is the way forward due to its simplicity
  and scalability
• IETF has standardised PHBs but has not mandated
  particular mechanisms to realise them
• Focus of this work -gt AF PHB
• AF PHB has outlined 4 classes and 3 drop
  precedence levels within each class
• RFC 2957 has recommended AQM to realise the drop
  precedence levels (to achieve differentiated
  dropping)
• Transactional data applications (interactive) are
  an important class of apps requiring low delay
• Currently available solutions provide only
  throughput guarantees
• Support available only for qualitative
  differentiation of delay and no quantitative
  guarantees are offered

4
Design Objectives

• The challenge is to build an AQM scheme that can
• Prevent Global Synchronisation
• Capture the dynamics of the underlying system
  accurately in real time thereby bypassing the
  parameter tuning problem
• Avoid the conservative approach of mathematical
  modelling
• Achieve the performance objective (regulate the
  delay below the control target)

5
Extensions to the PAQMAN Algorithm

• Extended algorithm called PAQMAN-DS
• Supports 3 drop precedences per queue
• Uses same queue threshold for all 3 precedences
• Uses different packet drop probability
  computation function for each precedence
• Employs coupled queue lengths

6
PAQMAN-DS Algorithm

• Input - average queue size observations for all
  three precedences over the past interval
• for every PI seconds
• Compute average queue size for each precedence
  over the past interval
• Predict average queue size for each precedence
  over the next interval
• Compute packet drop probability (PDP) for each
  precedence

7
PAQMAN-DS Flowchart
8
Action on each packet arrival
9
Salient features of the proposed Approach

• Bypasses parameter tuning by employing Data
  driven Adaptive Learning Approach
• Does not need any prior knowledge of the traffic
  model (no assumptions)
• Adapts well to changes in underlying traffic
• RLS algorithm converges in 2N iterations
• Tested with different types of traffic
• Achieves performance objective by regulating the
  delay below the control target

10
Performance Evaluation of PAQMAN-DS

• Two rate three colour marker used at edge
• Target delay at CR 50ms (hence, q037.5 pkts)
• Tested under two Scenarios involving long lived
  TCP and ON/OFF TCP for different subscription
  rates varying from 25 to 125 in steps of 25

11
Performance Metrics

• Average delay
• S.D of instantaneous queue size
• Link Utilization
• Green, Yellow and RED Packet Loss Ratios

12
Results
Performance metrics v/s subscription ratio - long
lived traffic
Performance metrics v/s subscription ratio
On/Off traffic
Target delay 0.05s
13
Summary of the Results

• Irrespective of the traffic load, PAQMAN-DS
  regulates delay below the control target as
  opposed to the other two approaches
• S.D. of PAQMAN-DS queue is low and steady in
  comparison to that of RIO and ARIO indicating
  that it is capable of minimizing delay variation
• It also achieves link utilization very similar to
  the other two approaches
• Leniency of RIO and ARIO result in lower Green
  packets being lost as compared to PAQMAN-DS

14
PAQMAN-DS Summary

• Simple
• Lightweight
• Does not make assumptions about the traffic model
• Capable of providing sustained and consistent QoS
  by regulating the delay below the control target
• Capable of discriminating in favour of IN
  contract traffic

15
Future Work

• Performance evaluation in the presence of more
  complex traffic mixes
• Performance evaluation in the presence of
  multiple bottleneck links