An Overlay MAC Layer for 802'11 Networks

1
An Overlay MAC Layer for 802.11 Networks

• Ananth Rao, Ion Stoica
• OASIS Retreat, Jun 2004

2
Motivation

• 802.11 hardware provides initial ease of
  deployability for many applications
• Mesh networks
• Long haul links
• Large Infrastructure Networks
• Are these apps stretching 802.11 beyond its
  design goals (Wireless LANs)?

Internet Gateway
3
Problem 1 Different Data Rates
R
B
A
4
Problem 2 Unpredictability
2
1
3
4
5
5
Problem 3 Forwarding on Behalf of Others
Ethernet
1/3
1/2
1/6
1/9
1/6
1/6
1/9
1/9
This problem cannot be solved by local scheduling
or queue management algorithms like WFQ
6
Case for an Overlay MAC (OML)

• Several new MACs have been proposed to solve
  these problems
• We try and leverage some of these techniques in
  OML
• Hardware vs. Software MACs
• Huge cost advantage over building a new MAC
• Flexible and easy to implement application
  specific policies
• Facilitates experimental research on MAC layer
  issues
• Accurate timing not possible at the software
  level
• Devices dont expose all information (eg. cannot
  carrier-sense and obtain result)
• Cannot change the physical layer (eg. spread
  spectrum techniques)

7
Constraints

• Do not want to modify the MAC protocol
• We have no control over a packet once it is in
  the cards buffer
• It may be queued behind another packet
• Solution Disable buffering in the lower layers
• There may be some back-off and retransmission
• Solution Ensure that with high probability,
  there will be no back off or retransmissions

Use reservation-based time slots to schedule
transmissions
8
Time Slots

• Assume local synchronization of clocks
• Use coarse-grained (compared to packet
  transmission times) time slots
• 20ms slots, 2ms to transmit an MTU sized packet
• Clocks are synchronized by estimating the 1-way
  delay of packets
• Assume back-off is negligible and there are no
  retransmissions at the MAC layer
• Use the minimum delay from the previous 20
  packets to reduce errors

9
Weighted Slot Allocation

• Algorithm to allocate time-slots to competing
  nodes
• Only requires knowledge of the IDs of other
  competing nodes no additional signaling
• Uses a pseudo-random hash function
• For ease of explanation
• Stage 1 Assume the diameter of the network is
  one
• Stage 2 Multi-hop networks of larger diameter

10
WSA Diameter One Network

• Each node keeps track of an active_list, the set
  of nodes with active flows
• Each packet includes the current queue length in
  the OML header
• Use a pseudorandom hash function to decide the
  winner
• The node with the highest Hi is allowed to
  transmit

11
WSA Multi-hop Network

• Use the same hash-based mechanism
• active_list only contains other nodes that
  compete with a given node
• Consider only the hash values of competing nodes
  in deciding the winner
• Assume that nodes interfere only up to k-hops
• k is a tradeoff between losing possible channel
  reuse opportunities, and using the underlying MAC
  to resolve contention

12
Partially Overlapping Contention Regions
Contention region of j
Hi gt Hj gt Hk

• 1-hop neighbors of the winner initiate an
  inactivity timer to detect when this happens
• If i is more than 1-hop away from j, it is
  notified using a control message

i
j
k
Contention region of k
13
Persistent Allocation of Slots
Time

• Slots maybe
• Available for contention
• Assigned to a particular node
• If the nodes queue goes empty, the rest of the
  slot is used by the node with the next highest
  hash

1
2
3
4
7
8
6
5
0 ms
24 ms
48 ms
72 ms
Groups of 8 slots each of length 3ms
14
Amortize the Cost of Contention Resolution

• Nodes that transmitted successfully in the
  previous slot with index i own the slot with
  probability (1-p)
• Cost is amortized because
• A time-slot is much longer than a packet
  transmission
• Nodes compete for an average of 1/p slots at a
  time
• Orthogonal to method used to resolve contention
  for a slot

Time
1
2
3
4
7
8
6
5
C
C
C
0 ms
C
C
24 ms
48 ms
72 ms
15
More Details

• The protocol is optimistic in assigning a slot
  to a node
• If a slot is assigned to more than 1 node, 802.11
  will still do its best to resolve contention
• If this happens very often, resource allocation
  goals may not be met
• Queues vs. Flows
• Because of TCP congestion control, queues become
  empty and full very often
• If the queue is empty, a node will signal that
  other nodes may send in that slot

16
Simulation Results

• Qualnet Network Simulator
• Commercial software from www.scalable-networks.com
  
• Packet level simulator similar to ns2, but faster
  and more scalable
• Models collisions, interference and contention
• Use 802.11a at 54 Mbps
• 20 slots of 20 ms each, p0.05

17
Fairness in a Multi-hop network
18
Flexibility in resource allocation
Different allocation policies can result in very
different outcomes hence the MAC layer must be
flexible
Qualitative differentiation is possible by
assigning different weights to flows
19
Test-bed

• Hardware
• ASUS Pundit barebones system
• Celeron 2.4 Ghz, 256 MB
• Netgear WAG511, 802.11a

• Software
• RH 9.0, Kernel 2.4.22
• Madwifi driver for Atheros
• Click modular router

20
Click Architecture
Push
1
Pull
FromDevice
DecapTimestamp
ToDevice
ContentionResolver
TimeslotEnforcer
EncapTimestamp
Rest of the Router
1
21
Results (Test-bed Data Rates)
Equal throughput default 802.11
Equal channel access time - OML
OML can achieve fairness in terms of transmission
time, or in terms of throughput
22
Results (Testbed Fairness, Flexibility)

• In 26 of the experiments, one flow was
  completely shut out by the other
• k2 leads to underutilization of the channel in
  some cases

When all nodes can hear each other, WSA
guarantees weighted fairness
23
Results (Test-bed)
C
A
B
24
Conclusions and Future Work

• Coarse-grained scheduling on top of 802.11 is a
  very powerful technique to
• alleviate inefficiencies of the MAC protocol in
  resolving contention
• overcome the lack of flexibility of assigning
  priorities to senders
• Future work
• Understand performance problems better though
  more measurements on the test-bed
• More benchmarks of OML