Bluetooth: Scheduling

1
Bluetooth Scheduling

• Reference Enhancing performance of asynchronous
  data traffic over the Bluetooth wireless ad-hoc
  network Das, A. Ghose, A. Razdan, A. Saran,
  H. Shorey, R. Proceedings, IEEE INFOCOM 2001,
  pp. 591 -600 (BTSche-2.pdf)

2
Introduction

• Contribution
• Propose two SAR policies with the aim of
  increasing link utilization and decreasing
  end-to-end delay of data packets
• MAC scheduling algorithm are needed to achieve
  fair sharing of bandwidth, high link utilization
  and low queue occupancy
• Demonstrate that Round-Robin scheduling is unable
  to meet these requirements and proposed three new
  scheduling algorithms
• Investigate the performance improvement provided
  by FEC and ARQ schemes
• Compare the performance of different versions of
  TCP over Bluetooth

3
Bluetooth Protocol Stack
4
Design Issues in BT

• 1. Segmentation and Reassembly schemes
• Supporting a maximum transmission unit (MTU) size
  larger than the largest baseband packet
• Note that the payload size (without FEC)
• 5 slot packet 339 bytes (67.8 bytes/slot)
• 3 slot packet 183 bytes (61 bytes/slot)
• 1 slot packet 27 bytes
• Define slot_limit as the maximum of slots
  across which a baseband packet can be sent
• May be less than 5 due to presence of SCO
  connections or due to very high bit error rate in
  the wireless channel
• This parameter can be conveyed by the LMP to the
  L2CAP through a signaling packet

5
Design Issues in BT (cont)

• Two SAR schemes
• SAR-Best Fit (BF)
• SAR-Optimum Slot Utilization (OSU)

6
Design Issues in BT (cont)

• Scheduling algorithms in BT
• Master-driven Round-Robin scheduling
• Achieve fair sharing of bandwidth and high link
  utilization when each such connection has equal
  data flow
• However, each slave in the piconet has varying
  data input rates
• Consequently, numerous baseband slots are wasted
  by polling sources with low input rate, thereby
  decreasing link utilization, increasing queuing
  delay and leading to unfair sharing bandwidth

7
Design Issues in BT (cont)

• Two basic methods for scheduling
• 1. Queue Priority based on Flow Bit
• Assign priority to per-slave baseband queues at
  the master (similar queues at the slave) based on
  the pending data in the L2CAP buffers, and use
  the flow bit field for this purpose
• The flow bit is set when the number of packets in
  the L2CAP buffer for a particular slave is larger
  than a threshold buf_thresh
• At the master, variable flow to quantify the
  traffic rate on the wireless channel, which is
  set when the flow bits for packets traveling in
  either direction is turned on

8
(No Transcript)
9
Design Issues in BT (cont)

• 2. Queue Stickness
• To reduce mean queue occupancy, propose to
  transmit a number of baseband packets
  successively (quantified by a parameter
  num_sticky) for each queue having the flow
  parameter set
• Three scheduling algorithms
• AFP, Sticky, StickyAFP

10
Design Issues in BT (cont)

• 1. Adaptive Flow-based Polling (AFP)
• Define P0 the initial polling interval
• AFP uses an adaptive polling interval P, whose
  value is changed based on the traffic rate
• 1. If flow 1 and the HOL packet is a data
  packet, transmit the data packet and set the
  polling interval P to P0 (High rate for this
  slave)
• 2. If flow 0 and the HOL is a data packet,
  transmit the data packet and keep the polling
  interval unchanged
• 3. If a polling packet is transmitted and a null
  packet is received, double the current polling
  interval P unless a threshold value Pthresh is
  reached

11
Design Issues in BT (cont)

• 2. Sticky
• 1. If flow 1, a maximum of num_sticky packets
  are transmitted for that queue
• 2. If flow 0, one packet is transmitted for
  that queue, as in Round-Robin scheduling
• 3. Sticky AFP
• Similar to AFP except that when flow 1 and the
  HOL packet is a data packet, a maximum of
  num_sticky packets are transmitted for that queue

12
Design Issues in BT (cont)

• Error Handling
• For ARQ scheme, quantify timeout by a maximum
  number of retransmissions tx_thresh
• Using two state Markov channel model
• Study the effect of using FEC on the baseband
  payload and of varying the parameter tx_thresh in
  the ARQ scheme on the data throughput
• Propose Channel State Dependent Packet (CSDP)
  scheduling
• Upon encountering a packet loss (NACK), CSDP
  policies defer the retransmissions to that slave
  till the next polling instant

13
Design Issues in BT (cont)

• Number SCO connections
• The master sends SCO packets at regular
  intervals, the so-called SCO intervals TSCO
  (counted in slots)
• In the presence of one SCO link, 1 and 3 slot ACL
  packets can be supported
• TCP variants
• TCP Tahoe, Reno, New Reno, Sack

14
Simulation Model
15
Simulation Model (cont)

• Performance Metrics
• Throughput, end-to-end delay, link utilization
• Fading Channel Model

rB 55.8 ms
rG 437.5 ms
1.263 x 10-3
6.879 x 10-5
16
Simulation Results

• Performance evaluation SAR schemes
• Use Round-Robin scheduling at the MAC level
• Assume an error-free channel
• Compare SAR-BF vs. SAR-OSU
• Figures 5 7 for slave 1

17
Simulation Results (cont)
For slave 1
18
Simulation Results (cont)
For slave 1
19
Simulation Results (cont)
20
Simulation Results (cont)

• Observations
• 1. Higher throughput, higher overall link
  utilization, and lower end-to-end delays can be
  obtained by using SAR-OSU over SAR-BF
• 2. The frequent fluctuations are due to the
  bursty sources (simulated by intermittent CBR
  traffic)
• 3. Since two TCP connections are active between
  10s 20s, fair sharing of bandwidth leads to a
  drop in the individual throughput of TCP
  connection to slave 1, while the overall link
  utilization is seen to increase

21
Simulation Results (cont)

• L2CAP Buffer size
• Use Round-Robin scheduling and assume an
  error-free channel
• Figures 8 9
• Observations
• 1. The average TCP throughput becomes almost
  constant for a buffer size greater than four for
  the persistent TCP connection
• 2. Conclude that a buffer size of four to six
  will optimally satisfy the memory requirements of
  a generic BT device (set L2CAP buffer size 5 )

22
Simulation Results (cont)
Long TCP slave 1 Short TCP slave 2
23
Simulation Results (cont)
24
Simulation Results (cont)

• Scheduling Algorithms
• Figure 10
• AFP and Sticky algorithms give significantly
  improved performance compared to RR
• The throughput of Sticky increases with increase
  in the value of num_sticky and is approximately
  the same as AFP for num_sticky16
• Figure 11
• The throughput of StickyAFP with num_sticky16 is
  better than that of AFP and StickyAFP with
  num_sticky4, but not very significant
• Figure 12
• High link utilization is obtained for StickyAFP
  (num_sticky16), Sticky (num_sticky16) and AFP,
  as compared to Round-Robin

25
Simulation Results (cont)
AFP, Sticky(16)
Sticky(4)
Sticky(2)
RR
For slave 1
26
Simulation Results (cont)
StickyAFP (16)
AFP, StickyAFP(4)
For slave 1
27
Simulation Results (cont)
StickyAFP(16)
AFP
Sticky(16)
RR
28
Simulation Results (cont)

• Figure 13
• The Sticky algorithm is found to have the lowest
  end-to-end delay while StickyAFP has the highest
• 1. By increasing the polling interval, AFP
  decreases the number of poll packets (for those
  queues that have less data) which otherwise cause
  underutilization of available bandwidth, and
  hence increases link utilization
• 2. Sticky reduces queue occupancy by transmitting
  multiple packets consecutively from queues with a
  high backlog, hence preventing queue overflow and
  reducing end-to-end delay

29
Simulation Results (cont)

• 3. StickyAFP causes a marked increase in the
  end-to-end delay of intermittent CBR traffic
  because flow is set infrequently for such bursty
  sources
• Additionally, each cycle has a larger duration
  due to other slaves being served num_sticky times
• Authors infer that AFP and Sticky(16) result in
  the best overall performance

30
Simulation Results (cont)
StickyAFP(16)
RR
AFP
Sticky(16)
31
Simulation Results (cont)

• Effect of Error Correction Schemes
• For different values of tx_thresh (max. of
  retransmissions of baseband packets)
• AFP vs. CSDP-AFP (w/ and w/o FEC)
• Observations (Figures 14, 15, 16)
• Performance degradation in the presence of errors
• Additional reduction in link utilization and
  increase in end-to-end delay due to the use of
  FEC
• When FEC is added, the performance is independent
  of tx_thresh, and hence of the ARQ scheme

32
Simulation Results (cont)

• ARQ leads to efficient error recovery and for
  values of tx_thresh gt 4, the performance does not
  vary significantly ? set tx_thresh 5
• Figure 16, CSDP versions of the proposed
  scheduling algorithms do not give a significant
  performance improvement and their relative
  performance is the same as that in the error-free
  channel condition
• CSDP versions do not improve the performance in
  the presence of a link level ARQ scheme

33
Simulation Results (cont)
AFP (error-free channel)
CSDP-AFP, AFP
CSDP-AFP (with FEC)
34
Simulation Results (cont)
CSDP-AFP (with FEC)
CSDP-AFP, AFP
AFP (error-free channel)
35
Simulation Results (cont)
CSDP-StickyAFP(16)
CSDP-AFP, AFP
CSDP-Sticky(16)
36
Simulation Results (cont)

• Varying voice connections
• Using AFP
• Figures 17 18
• Observations
• The throughput decreases and end-to-end delay
  increases as the number of SCO connections
  increase
• Higher throughput and lower end-to-to-end delay
  is obtained for slot_limit 3 than for
  slot_limit 1

37
Simulation Results (cont)
NSCO0, slot_limit5
TSCO6, NSCO1, slot_limit3
38
Simulation Results (cont)
TSCO6, NSCO2, slot_limit1
TSCO4, NSCO1, slot_limit1
TSCO6, NSCO1, slot_limit1
TSCO6, NSCO1, slot_limit3
NSCO0, slot_limit5
39
Simulation Results (cont)

• TCP variants (Fig. 19)
• The difference in throughput is insignificant
  which clearly illustrates that the efficient link
  layer ARQ scheme of BT eliminates the need for
  modifications at the transport layer for error
  recovery

40
Simulation Results (cont)