Adaptive Inverse Multiplexing for Wide-Area Wireless Networks

1
AdaptiveInverse Multiplexingfor Wide-Area
Wireless Networks

• Alex C. Snoeren
• MIT Laboratory for Computer Science
• IEEE Globecom 99
• Rio de Janeiro, December 5, 1999

2
context

• Goal Provide speech and graphical interfaces to
  wireless devices over wide-area networks
• Challenge Construct a well-behaved high
  bandwidth channel out of low bandwidth shared
  access technologies

3
inverse multiplexing

• Idea simulate a large logical channel out of
  some number (called a bundle) of smaller ones

Inverse Multiplexor
Inverse Multiplexor
High Bandwidth Link
High Bandwidth Link
Low Bandwidth Links
4
goals

• High link utilization and low fragmentation
• Low bandwidth wireless links
• Tight reordering constraints
• TCP doesnt handle reordered packets well
• Adaptive scheduling
• Throughput of shared wireless links is unstable
  over many time scales

5
contributions

• Standard Inverse Multiplexing
• Commonly used in ISDN, fractional T1/T3, ATM
• Private links with no contention
• Stable similar channel characteristics
• Link Quality Balancing
• Adapts to varying capacity shared access channels
• Efficient bandwidth utilization
• TCP-friendly reordering bound

6
outline

• Scheduling techniques
• Link Quality Balancing with stable links
• Adaptation
• Measuring and reacting to channel variations
• Implementation results
• Constant Bit Rate (CBR) Traffic
• TCP flows

7
known scheduling methods

• Round Robin
• Does not assure optimum link usage
• Provides no bounds on delay, ordering
• Deficit Round Robin, Fair Queuing
• Provide efficient link usage, but...
• Require information about queue lengths
• In CDPD, queues are often buried inside the
  networks, hence information is unavailable
• Dont provide ordering guarantees

8
deficit round robin
2
1
6
5
7
3
4
8
Inverse Multiplexor
9
fragmentation an extreme
10
weighting
x2
x1
x2
x1
11
link quality balancing

• Idea Fragment traffic in proportion to
  individual link throughputs
• For each link, compute a relative MTU
• For fastest link, use optimum MTU
• On all other links, use a proportionately smaller
  one
• Fragment packets to fill MTU-sized buckets
• Last fragment arrival times are the same on each
  link
• Guarantees no inter-round reordering only
  possible reordering occurs in the same round
• Requires no information on queue lengths
• Work conserving provides maximal link usage

12
our approach balancing
x2
x1
8
Inverse Multiplexor
x2
x1
8
Inverse Multiplexor
13
measurement

• Problem Individual link throughputs are highly
  variable over many time scales
• How do we measure current throughput?
• Absolute values are difficult and expensive to
  obtain
• Without synthetic traffic, we are limited by the
  offered load who knows if it actually is
  driving the links to full capacity
• Synthetic probes are problematic
• Without priority queuing, introducing synthetic
  traffic may cause loss of actual traffic

14
link quality metric

• Solution Dont! Relative metrics suffice
• Simply maintain proportional estimates
• End-to-end bandwidth probing will do the rest
• But which metric?
• Packet arrival times
• Theoretically ideal, but far too noisy to be used
  in reality
• Short-term throughput
• Similarly difficult to measure
• Loss Rates
• With bounded queues, loss rates are a rough
  indicator of appropriate throughput, and easy to
  measure

15
feedback loop

• Invariant Always schedule traffic so that
  quality metric will be identical across links
• As a corollary, any perceived deviation at the
  receiver implies an improper estimate
• Use the receivers data to periodically update
  the Multiplexors scheduling proportions
• End-to-end bandwidth probing should cause the
  weakest link to fail first and/or more often
• Links are asymmetric measure both ways

16
cbr traffic
Throughput (bits/sec)
Time(secs)
17
tcp traffic
Throughput (bits/sec)
Time(secs)
18
evaluation future work

• LQB handles shared wireless links well
• Fragmentation is minimal
• Reordering is tightly bounded
• Adapts well to varying channel characteristics
• But wed like to find a better metric
• Loss rates are delayed and very coarse grained
• Perhaps filtering functions exist for
  inter-packet arrival times