An Adaptive Nonpreemptive Scheduling Framework for Delay Bounded Traffic in Cellular Networks

1
An Adaptive Non-preemptive Scheduling Framework
for Delay Bounded Traffic in Cellular Networks

• Yaser Khamayseh
• Department of Computing Science
• University of Alberta

2
Agenda

• Introduction
• System model
• Problem Statement
• Related work
• System Model
• Delay Bounded Adaptive Scheduling Framework
  (DBAS)
• Two Channel Packing Algorithm
• Simulation Results
• Conclusion and Future Work
• Questions

3
Introduction

• Multimedia streaming flows
• High data rates
• Delay sensitive
• Wireless networks
• Limited resource
• Call admission control and scheduling algorithms
• Application and transport layers

4
Some Related Work

• Many existing results diverse formulations
• Main QoS tools
• Schedulers short term decisions
• Call Admission Control (CAC) long term decision

5

• Joshi et al. MobiCom 2000
• Downlink scheduling in CDMA data networks
  
• Evaluated many scheduling algorithms on traces of
  HTTP traffic
• Scheduler exploits job size, and user channel
  condition
• However, a user channel is assumed to remain
  constant during a connection's lifetime

6

• L. Xu, X. Shen, and J.W. Mark Wireless Com.
  '02
• Dynamic bandwidth allocation with fair
  scheduling for WCDMA systems
• Considered heterogeneous mix of traffic on the
  uplink e.g., 10 voice, 4 best effort, and 1
  video flows
• Devised a slot-by-slot weighted fair scheduler
  for allocating bandwidth (taking MAI into
  account)
• Obtained deterministic delay bounds on a
  session's flow if the flow conforms to the
  parameters of a leaky bucket regulator, and the
  cell load allows the base station to allocate
  bandwidth more than the token arrival rate
  parameter
• Results effective sharing of the cell's soft
  capacity
• However, a scheduler alone cannot provide
  assurance of continued QoS as users roam in the
  cell.

7

• Kwon, Choi, Bisdikian, and Naghshineh Wireless
  Networks '03
• QoS provisioning in wireless/mobile
  multimedia networks using a adaptive framework
• Considered a rate adaptive framework (Rmin,
  Rsat, Rmax)
• Unbounded adaptations during a flow's lifetime
• Fixed cell capacity environment
• Cell capacity max. of users that can be served
  at Rsat
• Connection duration exponential with mean 1/?

8

• Cell i residence time exponential with mean 1/
  h (i.e., h is the handoff rate
• After cell residency user moves to an adjacent
  cell with some probability.
• Cell Overload Probability
• tpredict is a CAC parameter
• Given a distribution l, n, r of users at time t
• P(l,n,r),OL Prob target cell has maximum load
  at
• time t tpredict
• Admission policy P(l,n,r),OL Padmit
• P(l,n,r),OL is approximated by the tail of a
  Gaussian distribution.

9

• F. Yu, V. Wong, and V. Leung INFOCOM 2004
• A new QoS provisioning method for adaptive
  multimedia in cellular wireless networks
• Motivated by the ability to change rate
  dynamically in UMTS
• Fixed cell capacity environment
• Considered K service classes, each class has its
  own set of possible service rates, and reward
  function(s) (to reflect the gain in providing a
  service), e.g.
• class 1 (128, 192, 256) Kbps, and
• class 2 (64, 96, 128) Kbps
• Rate adaptation occurs only on call arrival, and
  departure

10

• Ask for a policy (combined CAC and bandwidth
  allocation) that is optimal on the long term
• Formalized the problem as a semi-Markov decision
  process (SMDP)
• System state includes the number of users in
  each class and data rate, in the target cell and
  all neighbouring cells.
• Used an average reward Reinforcement Learning
  (RL) solution method to handle the problem.

11
System Model

• GSM-like cellular system
• C channels
• Fixed satisfactory data rate for each user
• Transmission time between the base station and
  streaming server is negligible

12
BSC
Data IP Network
SGSN
GGSN
BSC
Multimedia Server
Initialization
Inter Cell Handoff
Media Flow
Service interruption delay
Termination
Start of service delay
Base Station
User Terminal
Server
13
Problem Formulation (Non-preemptive Scheduling)

• Adopted framework similar to scheduling tasks
  with release times, and due dates i.e., each
  connection request ri has
• li connection length
• ai arrival time
• di maximum tolerable delay
• aggregates start of service delay, and service
  interruption delays
• ri must start during the interval ai, aidi
• wi weight (models connection priority at some
  instant)
• Adjusted for handoff connections and/or preferred
  connections

14

• Tplan (planning horizon) determines requests
  eligible for receiving service
• The Ideal non-preemptive scheduler
• Selects a subset R of connections of maximum
  total weight.
• Challenges Limitations
• Hard optimization problem if Tplan, connection
  lengths, and tolerable delay assume arbitrary
  large values.
• When Tplan lt 1 (all selected requests are
  eligible to receive service), a simple priority
  based scheduler (PBS) is optimum.
• In contrast, if Tplan gt 1, no PBS is optimum.

15
Proposed Adaptive Framework

• Has two components a front end back end
• Back End
• Uses a scheduling heuristic the 2-Channel
  Packing (2-CP) algorithm
• Works with any choice of Tplan
• Optimum in simple cases (e.g., Tplan 1)
• Can achieve more than 28 improvement over a 1-CP
  heuristic

16

• Based on dynamic programming

17
Delay Bounded Adaptive Scheduling Framework (DBAS)

• Front end component
• Estimates Tplan in case of backlog
• Calls the scheduler (part of the back end)

18
DBAS Back End

• let S the subset of requests whose remaining
  tolerable delay lt Tplan.
• for i 0, 1, .. C/2-1
• invoke function 2-CPS to schedule requests to be
  served by channels 2i, and 2i1 remove the
  scheduled jobs from S.
• if (2(i1) lt C) obtain an optimum packing for
  channel C
• start executing the scheduled requests

19
Two-Channel Packing Scheduler

• -
• is the set of all undelayed and delayed
  instances of traffic requests. Delayed instances
  are bounded by Tplan
• - nDR is the number requests in DR
• L1,2, , nDR is a list of the traffic
  instances in DR sorted in nondecreasing order of
  their starting times.
• Lk refers to the identity of request rj,d that
  belongs the list L.

20
2CPS Dynamic Program

• S(k)i,j the first k entries of L is of type
  S(k)i,j if
• no two instances in the subset correspond to the
  same request in R
• the subset can be scheduled over two channels
  such that the latest requests scheduled for
  service over the two channels start at instants
  i, and j, respectively, where i j.
• Note that, the special case where i j -1
  indicates that no request is scheduled for
  service over both channels.

21
2CPS Dynamic Program (Pseudo-code)

• 1. construct the set DR, and the ordered list L.
• 2. perform the following initializations
• case both channels are not currently serving
  requests
• set S(0)-1,-1 f break
• case one channel is empty, the other is
  currently
• serving a request r
• set S(0)-1, 0 r break
• case both channels are busy serving two
  requests r1 and r2
• set S(0)0, 0 r1, r2 break

22
2CPS Dynamic Program (Pseudo-code)

• 3. for k 1, 2, . . . , nDR
• 3.1 for all relevant pairs of time instants (i,
  j), i j,
• i, j e 0, Tplan
• S(k)i, j maximum weighted subset of
• (S(k-1)i, j,
• S(k-1)i, j U Lk where Lk can be
  tightly packed with S(k-1)i, j to obtain
  a valid S(k)i, j set )
• 4. return a set from the collection S(nDR)., .
  with the
• maximum possible total weight

23
Trace Analysis

• Modified predictive CAC algorithm
• Predict the system overload probability (Pco) at
  the end of estimation time T.
• Accept if Pco lt Pqos
• Else reject

24
Simulation Results

• Two sets of performance studies
• Online System
• Trace Analysis with adaptive Tplan

25
Results (Trace Analysis)

• Forced Terminations-vs-Cell Load

• Expectation the modified (predictive) CAC is
  strong in minimizing forced terminations
• Note the 2-CP heuristic works with full
  knowledge of the traffic (Tplan is very large)
• Findings the heuristic indicates opportunities
  to achieve improvement.

26
Results (Trace Analysis) cont.

• Effective Throughput-vs-Cell Load

27
Results (Trace Analysis) cont.

• Completed-vs-Cell Load
• Blocked-vs-Cell Load

28
Results (Adaptive online)

• Comparison against a priority based scheduler
  (PBS)
• Request priority wi lI / di
• Favours longer requests (in case of a tie among
  two requests with equal tolerable delays)
• Effective Throughput-vs-Cell Load
• The adaptive framework delivers better results

29
Concluding Remarks

• The proposed framework combines
• use of effective scheduling heuristic
• efficient use of the heuristic by adapting the
  length of the planning horizon according to the
  offered load
• The preliminary results are encouraging
• Currently, considering extending the framework to
  provide preemptive scheduling.

30
(No Transcript)