Sigmoid Function Based Dynamic Threshold Scheme for SharedBuffer Switches

1
Sigmoid Function Based Dynamic Threshold Scheme
for Shared-Buffer Switches

• By
• Boran Gazi, Zabih Ghassemlooy
• Optical Communications Research Group
• Northumbria University, Newcastle

2
Outline

• Introduction
• Buffer Management in a Packet Switch
• Dynamic Thresholds
• Fuzzy Thresholds
• Simulation Model
• Results and Discussions
• Summary

3
1. Introduction

• Buffering is required to resolve contentions in a
  packet switch
• Shared Buffer Switches (SBS) are far better than
  other buffering techniques such as
  output-queueing, input-queueing, recirculation
  buffering etc.
• SBS are better-
• Provide low packet loss rates
• Buffer space is better utilised
• Provide flexibility when allocating buffer space
  for contending packets

Contd.
4
1. Introduction

• Shared Buffer Switches are prone to
• high packet loss rates
• unfair use of buffer space.
• Buffer management schemes are required to
  overcome these two problems
• In this work fuzzy thresholds has been utilised
  as a buffer management policy in SBS

5
2. Buffer Management in a Packet Switch

• There are two main categories of Buffer
  Management
• Static Policies is based on static parameters
  set based on statistical information and repeated
  simulations
• Dynamic Policies attempts to control the buffer
  space based on the information from environment
  variables

Contd.
6
2. Buffer Management in a Packet Switch

• Examples of Static Policies
• Complete Sharing,
• Complete Partitioning,
• Sharing with Minimum Allocation (M. Irland),
• Sharing with Maximum Queue Lengths,
• Hybrid.
• Examples of Dynamic Policies
• Dynamic Thresholds (Choudhury Hahne),
• Push-Out, Harmonic Buffer Management,
• Adaptive Control.
• Push-Out policy is naturally adaptive, but almost
  impossible to implement

7
3. Dynamic Thresholds (Choudhury Hahne)

• A single threshold for all queues
• Threshold T(t) is directly proportional to
  available buffer space at time t.
• Simply a packet is rejected if Qi(t) gtT(t)
• a has to be set manually according to traffic
  phase and/or switch characteristics

Optimal a ½ a 2
8
4. Fuzzy Threshold

• Packets are admitted or blocked by using a notion
  of fullness
• Employs sigmoid function to determine fullness
• Unlike DT, it employs multiple thresholds

Contd.
9
4. Fuzzy Threshold

• A packet is admitted to queue i at time t with a
  probability of 1- µi(t)
• ß is the share parameter (0 ltß 1)
• For very large a, admission policy is fixed
  rather than fuzzy

10
5. Simulation Model SBS Model

• Consists of an N x N switch
• Buffer space shared among N output ports
• Buffer Management unit determines queue
  thresholds, T(t), according to a policy
• Packet lengths are fixed (a packet lasts one
  frame-time)
• Each queue is served deterministically one
  packet per frame-time.

Contd.
11
5. Simulation Model Traffic Model

• Inputs of SBS are connected to N independent
  asynchronous traffic generators
• Traffic generators model Interrupted Poisson
  Process (IPP)
• ON-OFF durations are exponentially distributed
• Arrivals occur at ON durations with rate ?
• Traffic distributions can be symmetric and
  asymmetric (Hotspot)

12
6. Results and Discussions Optimal ß

• 32 x 32 switch
• 640-packets buffer space
• Input Load p 0.8
• Hotspot loads 0.8, 0.95 and 1.05
• Packet Loss Rate (PLR) performance for different
  hotspot loads
• Optimal ß
• 0.20 ß 0.25

Contd.
13
6. Results and Discussions Optimal ß

• No. of hotspots 3, 4 and 5
• Input Load p 0.8
• Packet Loss Rate performance for different number
  of hotspots
• Optimal ß
• 0.20 ß 0.25

Contd.
14
6. Results and Discussions DT versus Fuzzy
Threshold

• Hotspot load 0.95
• ß 0.25
• Dynamic Thresholds performs better
• Fuzzy Threshold employs stricter admission
  control for active queues
• More space is spared for inactive queues

PLR versus number of hotspot ports
Contd.
15
6. Results and Discussions Dynamic Thresholds
Queue length versus time

• DT spares a reasonable amount of buffer space for
  inactive queues
• Unused buffer space changes together with the
  activity of active queues

Contd.
16
6. Results and Discussions Fuzzy Thresholds

• Unused buffer space changes with traffic activity
  rather than active queues
• Fuzzy thresholds underutilise the spared space
  for inactive queues
• More hotspot packets are dropped

17
7. Summary

• Sigmoid Function Threshold (Fuzzy Threshold) uses
  the notion of fullness
• Achieves a reasonably good PLR performance
• Unlike DT, Employs multiple thresholds
• Threshold policy is rather strict
• Buffer space spared for inactive queues is
  underutilised
• A less strict policy should be employed to spare
  just enough space for inactive queues (i.e.
  ideally Push-Out policy)

18
Thank You!